diff --git "a/measurement-airoboros-l2-70b-gpt4-1.4.1.json" "b/measurement-airoboros-l2-70b-gpt4-1.4.1.json" new file mode 100644--- /dev/null +++ "b/measurement-airoboros-l2-70b-gpt4-1.4.1.json" @@ -0,0 +1,249736 @@ +{ + "measurement": [ + { + "key": "model.layers.0.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.013400986790657043, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.012927290052175522, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.008111074566841125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.00800967589020729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.007914060726761818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0064737857319414616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.026912858709692955, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.01268946472555399, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.007988039404153824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.007843940518796444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.008063625544309616, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.008470503613352776, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.007814689539372921, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.006717255804687738, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.00641877856105566, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.007115745451301336, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00639689713716507, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.006298080552369356, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006392695475369692, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00629305187612772, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006740371696650982, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006391115020960569, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006616147700697184, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00628971541300416, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.013400986790657043, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.011565840803086758, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.01092418935149908, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.005493824370205402, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0053792670369148254, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.00527279544621706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0025767649058252573, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.01446455530822277, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.010693219490349293, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.005424493458122015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.005174398422241211, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.005399877205491066, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.005730528384447098, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.005143448710441589, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.003184797242283821, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.002481648698449135, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0032321326434612274, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0024305328261107206, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0021736312191933393, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0024205571971833706, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0021580392494797707, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0022743318695575, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.002417908050119877, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0018926814664155245, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002149843145161867, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.011565840803086758, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.057023786008358, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04637880623340607, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.037798505276441574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03156445920467377, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02454913966357708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.018078919500112534, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03830276429653168, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03402973711490631, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.026443660259246826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.019625665619969368, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.019404470920562744, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019545016810297966, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016281327232718468, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.012142766267061234, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.010964127257466316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00977054238319397, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006826216354966164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00636114040389657, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005968777928501368, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005331955384463072, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005083473864942789, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005440987646579742, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0034722513519227505, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004108484834432602, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.057023786008358, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.012079847976565361, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.008976669982075691, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.007188574876636267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.006059103645384312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.004787835758179426, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0035554347559809685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.00751971872523427, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.006749738473445177, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.00541717978194356, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.004011338111013174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.003891248255968094, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.00380116468295455, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0035076553467661142, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.002737300004810095, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.002527901902794838, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0021363315172493458, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.002054515527561307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.001991602825000882, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.001973653445020318, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0018848503241315484, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0015027679037302732, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0019273804500699043, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0013294372474774718, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0017887209542095661, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.012079847976565361, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.06438535451889038, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.060016002506017685, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.058775413781404495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.05303316190838814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.027920624241232872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.026684634387493134, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.030890677124261856, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.028690382838249207, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.028249358758330345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.025335656479001045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.02423473447561264, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.015621278434991837, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.013882203958928585, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.013563740067183971, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.013489733450114727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.00792731624096632, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.007621843833476305, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.007601722609251738, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.007188989315181971, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.007173587568104267, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.004602627828717232, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.00544940447434783, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.004510657396167517, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.004537259694188833, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.06438535451889038, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.07407441735267639, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.0691351667046547, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.06763552874326706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.061059217900037766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.03203925862908363, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.030594637617468834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.03540310636162758, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.032903652638196945, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.032428182661533356, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.028997525572776794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.027669301256537437, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.017693545669317245, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.015492036938667297, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.015124424360692501, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.015039553865790367, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.008815260604023933, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.007844493724405766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.007817347534000874, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.007271421141922474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.007251252885907888, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.0046575311571359634, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.00484635541215539, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.004540734458714724, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.0033672393765300512, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.07407441735267639, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.03819708898663521, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.030490702018141747, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.024009520187973976, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.021203679963946342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.015423620119690895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.01122425775974989, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.024769753217697144, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.02251020446419716, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.016941599547863007, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.01326841302216053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.013116327114403248, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.012504289858043194, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.011042647995054722, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.008091648109257221, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.007243525702506304, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.0065151359885931015, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.005317335482686758, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.005005853250622749, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.005030014086514711, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.004626092035323381, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.003993740305304527, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.0047621759586036205, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.0032192848157137632, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.004130918998271227, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.03819708898663521, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.013755612075328827, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.010298789478838444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.005864711478352547, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.005705112591385841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.005432245321571827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0025321226567029953, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.011626857332885265, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.00962437316775322, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0062548196874558926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.004650006536394358, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.005089199170470238, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.005609977059066296, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.004549254663288593, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.002793955383822322, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.002201096387580037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0028266077861189842, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0017326120287179947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0015018533449620008, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0016172827454283834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0013149832375347614, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.001606361591257155, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0016039616893976927, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.001064094016328454, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0012463583843782544, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.013755612075328827, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.014939969405531883, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.010971572250127792, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.006421584635972977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.006126543506979942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.005751732736825943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0027404972352087498, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.012055481784045696, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.010082915425300598, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.006631952710449696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.004824809264391661, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.005176133010536432, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.005602680612355471, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.004687493667006493, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0028942900244146585, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.002285868627950549, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0028067142702639103, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0017019854858517647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0014596108812838793, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0015559166204184294, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0012251525186002254, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0015795424114912748, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.001536162686534226, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0010511103319004178, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.001126711955294013, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.014939969405531883, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07523959130048752, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.057815175503492355, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.049398645758628845, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0415172278881073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03191133216023445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02386663109064102, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04345681518316269, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0396430566906929, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03452671691775322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02393144927918911, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.023204896599054337, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.022326411679387093, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019050948321819305, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01543117593973875, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014464663341641426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011153063736855984, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008172184228897095, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0077918157912790775, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006661215331405401, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00601988285779953, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005813284311443567, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0058217658661305904, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004507139790803194, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0039070299826562405, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07523959130048752, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09134119749069214, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06210092082619667, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.048225484788417816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04025973752140999, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.037181612104177475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.025692421942949295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.054244108498096466, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04849221929907799, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04216238111257553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.025705311447381973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.026556994765996933, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02769872173666954, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02408502995967865, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019240599125623703, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.017926881089806557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.014489226043224335, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01216929778456688, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011812787503004074, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010585751384496689, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00989765115082264, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008933243341743946, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010395756922662258, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007598696276545525, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00903016421943903, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09134119749069214, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.07837946712970734, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.07387670129537582, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.07257820665836334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.06594417989253998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.03350868821144104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.03230910375714302, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.036773890256881714, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.03411875665187836, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.03381641209125519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.030770890414714813, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.029551999643445015, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.01847178488969803, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.016225192695856094, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.015969600528478622, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.015915242955088615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.009266272187232971, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.008636943995952606, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.008623353205621243, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.00816182792186737, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.008158819749951363, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.005162141285836697, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.005793672986328602, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.005085927899926901, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.004550590645521879, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.07837946712970734, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.09754252433776855, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.09216725826263428, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.09065655618906021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.08252513408660889, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.041799936443567276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.04034008830785751, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.04583039879798889, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.0425165593624115, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.042170483618974686, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.03844733536243439, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.036864642053842545, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.022814886644482613, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.019903216511011124, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.019604692235589027, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.019534563645720482, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.011342381127178669, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.010007600300014019, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.009989331476390362, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.009372668340802193, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.009370476938784122, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.005912867374718189, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.005936316214501858, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.005815788637846708, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.0038800491020083427, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.09754252433776855, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.08675035834312439, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.07898331433534622, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.07532230019569397, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.06918463855981827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.03757708519697189, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.03468551114201546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.04518641158938408, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.041367463767528534, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.038640737533569336, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.034127313643693924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.03329474851489067, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.02279055491089821, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.01979471743106842, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.018149785697460175, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.017744479700922966, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.011512796394526958, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.009960651397705078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.009801696054637432, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.00935293734073639, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.00913001038134098, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.006512295920401812, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.007032300811260939, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.00601436011493206, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.005498806480318308, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.08675035834312439, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.014618591405451298, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.012267633341252804, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.00981070939451456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.008386706002056599, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.00627940846607089, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0046131545677781105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.01005496084690094, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.009032872505486012, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.006616031751036644, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0051689427345991135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.005088103469461203, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0048549291677773, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.004248571116477251, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0030997744761407375, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.002767368918284774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.002452566521242261, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0017615273827686906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0016060513444244862, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0015734939370304346, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0013818326406180859, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0013567907735705376, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0014299890026450157, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0010168252047151327, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0010679904371500015, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.014618591405451298, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.013467859476804733, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.011508520692586899, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.00894941296428442, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.007676809560507536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.005752976052463055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.004133241251111031, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.009689346887171268, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.008698495104908943, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.006024597212672234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0048389434814453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.004777852445840836, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.00456572137773037, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.004054008983075619, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.002860552631318569, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0025097946636378765, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.002344077453017235, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0016629737801849842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0015001402935013175, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0015139564638957381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.001320952782407403, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0013718619011342525, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.001394552760757506, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0010618800297379494, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0010655829682946205, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.013467859476804733, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11011660099029541, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09690072387456894, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09201732277870178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07923592627048492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04932006075978279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.044213566929101944, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.058733511716127396, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.053458891808986664, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05078951641917229, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04107724502682686, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.038357120007276535, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02992418222129345, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.025459447875618935, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02357221022248268, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.023129189386963844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.014964539557695389, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012094573117792606, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011921148747205734, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010442087426781654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.010209315456449986, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007728317752480507, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00749287661164999, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006919396109879017, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004777991212904453, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11011660099029541, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13200624287128448, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10081067681312561, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0889894962310791, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07237827032804489, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05684700608253479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.044129930436611176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07238192856311798, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0657227486371994, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06110958755016327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04104398190975189, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03869885206222534, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03716018423438072, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.032167695462703705, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.028247080743312836, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.027251271530985832, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019133657217025757, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01624998450279236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01597936451435089, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013646120205521584, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013111625798046589, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01135975494980812, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012393233366310596, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010274537838995457, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01019685622304678, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13200624287128448, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.11293644458055496, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.10679198056459427, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.10506219416856766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.09597860276699066, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.04949721321463585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.04779047891497612, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.05453529208898544, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.05036134645342827, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.04993892088532448, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.045577522367239, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.04402731731534004, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.02753550559282303, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.024105040356516838, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.023743724450469017, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.023656830191612244, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.013847587630152702, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.012913329526782036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.012889890000224113, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.012225857004523277, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01222090981900692, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.007825355045497417, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.00873085018247366, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0077181109227240086, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.0069280993193387985, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.11293644458055496, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.13667632639408112, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.12937167286872864, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1273091435432434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.11641614139080048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0599554181098938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.05788889154791832, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.06594447791576385, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.061005499213933945, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.060497917234897614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.055187247693538666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.05318571627140045, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.033065393567085266, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.028714600950479507, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.02827843837440014, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.028178373351693153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.016457902267575264, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.014519226737320423, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.014491807669401169, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.013612879440188408, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.013607991859316826, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.008703500963747501, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.008731269277632236, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.008561388589441776, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.005848981440067291, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.13667632639408112, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.04921293258666992, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.04668398201465607, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.028258174657821655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.026349179446697235, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.02135397680103779, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.011830138973891735, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.048747673630714417, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.041573576629161835, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.02161886729300022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.02028924785554409, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.02203189581632614, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.02160262130200863, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.01854688674211502, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.010660359635949135, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.00756386574357748, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.010382684879004955, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.005842040292918682, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.00346121727488935, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.005656967405229807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0031294897198677063, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.00537630682811141, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.005339957773685455, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.0024847970344126225, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.002220517722889781, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.04921293258666992, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.02469894289970398, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.02197093516588211, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.020374488085508347, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.017842380329966545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.01103502232581377, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.009721755981445312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.014724674634635448, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.012749993242323399, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.011348681524395943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.00940489862114191, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.00894902367144823, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.007433102931827307, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.006082927342504263, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.005330168642103672, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.005139919463545084, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0037468180526047945, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0028568729758262634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.002769016195088625, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0025509458500891924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0024459625128656626, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0020184870809316635, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0019797233399003744, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0016912400024011731, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0014321935595944524, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.02469894289970398, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0266012791544199, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.023478496819734573, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.021732386201620102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.018938953056931496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.011707613244652748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.010249094106256962, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.014700816944241524, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.013552720658481121, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.012059764936566353, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.009903778322041035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.009373040869832039, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.007360122166574001, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.006422386504709721, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0056058494374156, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.005401432514190674, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.003687535645440221, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0029340432956814766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00283694826066494, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0025850122328847647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.002469512401148677, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0019452449632808566, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0019577890634536743, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0016789670335128903, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0013133406173437834, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0266012791544199, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12530039250850677, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11476241797208786, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1114124283194542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09905023127794266, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05701732262969017, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05324256420135498, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06436778604984283, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05933767557144165, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05803144723176956, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.049796346575021744, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04690178111195564, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03251999616622925, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028221316635608673, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.027142584323883057, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.026895839720964432, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016214333474636078, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013754921033978462, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013672552071511745, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012285000644624233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012206284329295158, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00830412469804287, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008078297600150108, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007890846580266953, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00491254311054945, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12530039250850677, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14146119356155396, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1193654015660286, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10385344922542572, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09000688046216965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0637824684381485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05078403279185295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08813130855560303, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08039939403533936, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06641644239425659, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.051762960851192474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04964544251561165, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04523259773850441, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03906260058283806, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03171538934111595, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029761051759123802, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.023095253854990005, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.018239304423332214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01737205684185028, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01621459424495697, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015144191682338715, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01332961767911911, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014298862777650356, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011203581467270851, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011407439596951008, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14146119356155396, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14067578315734863, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.13153208792209625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.12882263958454132, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1160285696387291, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.06290242820978165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.059995345771312714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.06963183730840683, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.06439526379108429, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.0636858195066452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.05651227757334709, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.053724855184555054, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.03509005904197693, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.030582157894968987, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.029937412589788437, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.029780985787510872, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.017528653144836426, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.015641026198863983, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.015599698759615421, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.014422367326915264, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.014402776956558228, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.00939891580492258, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.009811719879508018, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.009189524687826633, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.0070068808272480965, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14067578315734863, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1687474548816681, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1578361541032791, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.15461121499538422, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.13927249610424042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07548321038484573, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07200007140636444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08347244560718536, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07724884152412415, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.0764203816652298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06780445575714111, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06433388590812683, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04198801517486572, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03650892525911331, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03574411943554878, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03556372970342636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.020905354991555214, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.018233323469758034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.018183672800660133, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.016707463189959526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01668676547706127, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01090096216648817, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.010792098008096218, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.010649657808244228, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.006939759477972984, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1687474548816681, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.15999826788902283, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.14731614291667938, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1427636593580246, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.12967462837696075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.07116202265024185, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.06712157279253006, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.08119179308414459, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.07456278055906296, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.07248588651418686, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.06410986185073853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.061633240431547165, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04093390330672264, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.03535299375653267, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.03391628712415695, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.03356850892305374, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.020482996478676796, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.017726054415106773, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.017606567591428757, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.01639156974852085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.016204144805669785, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.011141461320221424, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.011234010569751263, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.010696913115680218, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.007940838113427162, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.15999826788902283, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.033202119171619415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.028638897463679314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.02608690969645977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.022671708837151527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.01473529264330864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.012551616877317429, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.019294871017336845, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.017393162474036217, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.015326932072639465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.012236198410391808, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.01170465350151062, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.00977549608796835, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.008309190161526203, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.007117711938917637, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.006811006926000118, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.004917903337627649, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0038016564212739468, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00366530055180192, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.003331255866214633, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0031543090008199215, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0026499105151742697, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0026508159935474396, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0022236243821680546, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0018782122060656548, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.033202119171619415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.03393050283193588, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0289459265768528, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.026024796068668365, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.022547798231244087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.014766046777367592, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.012389698065817356, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.019396930932998657, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.01783786714076996, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.015434768050909042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.012199175544083118, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.011637726798653603, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.009729844518005848, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.008442066609859467, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.007101848255842924, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.006750521715730429, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.004853717517107725, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0037337378598749638, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.003576848190277815, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0032422735821455717, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.003031965810805559, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0025497281458228827, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.002570410491898656, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0021007999312132597, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0017221390735358, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.03393050283193588, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1325659304857254, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1199989840388298, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11588037759065628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10228347778320312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.06061491370201111, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.055925145745277405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06885495036840439, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06359714269638062, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06195104494690895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05217806249856949, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04893889278173447, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03487623110413551, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.030320923775434494, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02894563600420952, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.028607327491044998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.017404338344931602, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01472142618149519, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014613907784223557, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012992252595722675, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012866649776697159, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008948914706707, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00876140408217907, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008457031100988388, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0054040164686739445, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1325659304857254, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1611240655183792, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12974298000335693, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11838266253471375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09855736792087555, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.071906678378582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.058881133794784546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08732090145349503, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07990213483572006, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07555852830410004, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.053357984870672226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05011313781142235, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.044928014278411865, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.038854580372571945, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03525806590914726, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03438859432935715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02278820239007473, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019611084833741188, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019344914704561234, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016358498483896255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015899095684289932, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013036265037953854, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014153668656945229, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011985570192337036, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01124913152307272, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1611240655183792, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16268208622932434, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.15211917459964752, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.14898352324962616, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1342530995607376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07391005009412766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07041216641664505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08169501274824142, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07568961381912231, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07484319806098938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06626494973897934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06289547681808472, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04131200909614563, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.036053165793418884, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.035283029079437256, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03509339317679405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.020633986219763756, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.01842242106795311, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.018375281244516373, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01694951206445694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01692289486527443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.010999595746397972, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.011533538810908794, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.010748024098575115, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008198333904147148, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16268208622932434, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19267740845680237, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18023857474327087, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1765848994255066, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1591428816318512, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08749895542860031, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08335301280021667, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09663074463605881, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.0895843431353569, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08862366527318954, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0784369707107544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07434357702732086, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04869631677865982, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.042448777705430984, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.041528210043907166, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04131157696247101, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0242585651576519, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.021126482635736465, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02106792852282524, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019304810091853142, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.019275253638625145, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012544947676360607, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012403403408825397, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012242665514349937, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.007814181968569756, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19267740845680237, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.17970271408557892, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1649450957775116, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1597602367401123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.14354455471038818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0801788792014122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.07549820095300674, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09118957072496414, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.08389904350042343, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08165367692708969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07161428034305573, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.06804773211479187, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04599051922559738, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.03977814316749573, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.038187917321920395, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.037812888622283936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02298332378268242, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.019882211461663246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.01975497603416443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.01825401559472084, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.018048975616693497, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.012395924888551235, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.012497162446379662, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.011911667883396149, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.008725009858608246, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.17970271408557892, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04140036553144455, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.03710227459669113, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.034825485199689865, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.03074968419969082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.018684083595871925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.016704218462109566, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.024183690547943115, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.021046457812190056, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.019196193665266037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.01605156995356083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.015332351438701153, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.012227265164256096, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.010050182230770588, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.008994614705443382, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.00872649997472763, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0061411806382238865, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0047406344674527645, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.004621841479092836, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.004233307670801878, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.004090454895049334, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.003285675775259733, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0031630212906748056, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00281543773598969, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0022075807210057974, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04140036553144455, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04197612777352333, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.037344664335250854, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.03466910868883133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.030490456148982048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.018647726625204086, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.01644977368414402, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.023338697850704193, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.021509874612092972, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.01921883411705494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.015950500965118408, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.015195942483842373, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.011736894026398659, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01021081767976284, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0089501291513443, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.008635961450636387, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0058822063729166985, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.004684577230364084, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.004540184512734413, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00416543148458004, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.003986875526607037, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.003091281047090888, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003127087838947773, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.002674353774636984, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0021114249248057604, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04197612777352333, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1438351422548294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1321907639503479, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.12830744683742523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.11446458846330643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.06575622409582138, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.06151708588004112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.07419289648532867, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06849101185798645, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06695465743541718, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05766332894563675, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.054456453770399094, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03748634457588196, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.032571665942668915, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.031349048018455505, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.031050221994519234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01870095357298851, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01588156819343567, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.015779200941324234, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.014210224151611328, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.014109691604971886, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009577164426445961, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009310307912528515, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.009134655818343163, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005642216186970472, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1438351422548294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1648535281419754, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13339148461818695, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12223637849092484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10512534528970718, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07278762757778168, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06024128571152687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08818034827709198, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08124067634344101, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07705572247505188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.055976543575525284, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05267203599214554, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04523245990276337, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03916988521814346, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03538242727518082, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.034440647810697556, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02273479290306568, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019220992922782898, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0189637690782547, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01615537703037262, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015627652406692505, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01257948949933052, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013429229147732258, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011427650228142738, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010201756842434406, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1648535281419754, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1684059351682663, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.15775728225708008, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1545887291431427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1399359107017517, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0770682841539383, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07348117977380753, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08522959053516388, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07892042398452759, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07803229242563248, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06928963214159012, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06607364118099213, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04318227618932724, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03762642294168472, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.036816973239183426, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03662390261888504, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.021567482501268387, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.01917615905404091, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.019123263657093048, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.017659295350313187, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01763102225959301, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.011487470008432865, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.011925957165658474, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.011226076632738113, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008393969386816025, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1684059351682663, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20149937272071838, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18883642554283142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18512201309204102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16764189302921295, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09220602363348007, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0879039540886879, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.101831816136837, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09435611218214035, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09333047270774841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08291394263505936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07893910259008408, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.051428958773612976, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0447804294526577, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04381828010082245, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.043591998517513275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.025617558509111404, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02229228802025318, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02222810499370098, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020410237833857536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02037862129509449, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013249383307993412, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013072887435555458, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012927504256367683, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008222782984375954, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18883642554283142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.18520313501358032, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.17000456154346466, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1647968739271164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.14747653901576996, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.08282958716154099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.07804904878139496, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09390570968389511, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.08636510372161865, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08430133014917374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07372210174798965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0698256865143776, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.047435540705919266, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.040953874588012695, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.03945009037852287, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.03909241035580635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.023713000118732452, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02053707093000412, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02041955664753914, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.018806034699082375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.018607312813401222, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.012823803350329399, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.012866049073636532, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.012366954237222672, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.008980617858469486, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.18520313501358032, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04583487659692764, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.04134560748934746, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.038757938891649246, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.03432672843337059, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.020776856690645218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.018612872809171677, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.026959514245390892, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.023557322099804878, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.021290011703968048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.017974019050598145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.01717527210712433, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.013650144450366497, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.011261451058089733, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.010026009753346443, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.009717102162539959, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.006846908945590258, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.005327604711055756, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.005184083245694637, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.004792498424649239, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.004625611938536167, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.003651685081422329, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0036134617403149605, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.003109842073172331, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.002574960235506296, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04583487659692764, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04332634061574936, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.03893408551812172, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.03585195168852806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03167062625288963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.01933753490447998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.016960956156253815, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.024713080376386642, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.022840404883027077, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.019845571368932724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.016711313277482986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.015974003821611404, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.012410642579197884, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.010829276405274868, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.009289710782468319, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.008896656334400177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.006209140177816153, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.004876678343862295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.004679192788898945, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.004369039088487625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.004137468058615923, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0032773748971521854, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0033050579950213432, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0027848491445183754, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0022164189722388983, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04332634061574936, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.15830770134925842, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.14592844247817993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1416969895362854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1266210973262787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.07270292937755585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.06812571734189987, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.08240123838186264, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.075792096555233, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0739460438489914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.06390205025672913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.060434434562921524, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.041690513491630554, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03603246808052063, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.03462982550263405, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.03430779278278351, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.020776085555553436, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01754090003669262, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.017419084906578064, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01572880521416664, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.015602526254951954, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.010637594386935234, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.010265948250889778, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.010095321573317051, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006181938108056784, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.15830770134925842, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18595418334007263, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1502102166414261, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13763219118118286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11258202791213989, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08214733749628067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06834769994020462, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09972865879535675, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09164517372846603, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0871853157877922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06140601262450218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0572010837495327, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05117811635136604, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.044310420751571655, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04002753645181656, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03896921128034592, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025802765041589737, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02184883877635002, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.021573608741164207, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018102124333381653, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017537172883749008, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014465000480413437, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01537391822785139, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01318602729588747, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011796537786722183, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18595418334007263, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1731298714876175, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1623917520046234, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1591598242521286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1443706899881363, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0795891135931015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07590436935424805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08796006441116333, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08151081204414368, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08056706190109253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0716857761144638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06839878857135773, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04458511620759964, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03886750712990761, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03802482783794403, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.037828244268894196, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.022251475602388382, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.019765648990869522, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.019708136096596718, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01821637526154518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.018186677247285843, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.011775961145758629, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012233065441250801, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.011504183523356915, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008537485264241695, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1731298714876175, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21016940474510193, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1972198784351349, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19335848093032837, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17541223764419556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0966457799077034, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09219871461391449, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10680907219648361, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09892550110816956, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09782721102237701, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0870526060461998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08304751664400101, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05403987690806389, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04701443016529083, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04599899426102638, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.045757684856653214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.026938682422041893, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0234668031334877, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.023398881778120995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.021527210250496864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.021490920335054398, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014004339464008808, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013856434263288975, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01367052923887968, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008850243873894215, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19335848093032837, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.19139911234378815, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.17602969706058502, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.17073294520378113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.15290768444538116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.08581041544675827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08098604530096054, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09714095294475555, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.08943113684654236, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08732689917087555, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07651901990175247, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07242094725370407, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0490606427192688, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04243205115199089, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.040884897112846375, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.040512822568416595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.024507537484169006, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.021271010860800743, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.021150648593902588, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.01950186677277088, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.01930355280637741, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013181999325752258, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013316291384398937, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.012695170938968658, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009280321188271046, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.19139911234378815, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0596616268157959, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.054673731327056885, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05236173793673515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04668978974223137, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.027265867218375206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.025130275636911392, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03281135857105255, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.02946614846587181, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.027782032266259193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02387213334441185, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.022687871009111404, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.016626408323645592, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.014060181565582752, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.013091626577079296, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.012856579385697842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00833533238619566, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.006852461490780115, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00674847885966301, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006189126055687666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0060768104158341885, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004417521879076958, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004414290655404329, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.003996927756816149, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0030738834757357836, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0596616268157959, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.054068423807621, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04931257665157318, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.046965356916189194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04169907048344612, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.024375630542635918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.022288046777248383, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.02890053205192089, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.026757411658763885, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.024855514988303185, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.021282434463500977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.020169468596577644, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.014507980085909367, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.012700841762125492, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011656858026981354, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011403397656977177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007256956771016121, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006032402627170086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0059118689969182014, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0054179104045033455, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005289008840918541, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0037930640392005444, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0038250384386628866, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003457997925579548, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0025251503102481365, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.054068423807621, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1697668582201004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1576848179101944, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1537165343761444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.13791310787200928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.07812873274087906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.07366267591714859, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.08788589388132095, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08086226880550385, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.07928687334060669, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.06927874684333801, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.06569080054759979, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04442151263356209, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.038456112146377563, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.03722410649061203, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.03692842647433281, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.022149844095110893, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.018882153555750847, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.018781503662467003, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.017058126628398895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01697292923927307, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.011372048407793045, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.011057908646762371, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.010898091830313206, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00676465081050992, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1697668582201004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17937570810317993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14921247959136963, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13794280588626862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11596529185771942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07985196262598038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06769667565822601, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09667201340198517, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08912937343120575, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08398135006427765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06275630742311478, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.057032570242881775, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04948838800191879, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.042825181037187576, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03865894675254822, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03762959688901901, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024771692231297493, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020719919353723526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02039005421102047, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017514145001769066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.016964731737971306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01346397865563631, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01416557002812624, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01216246746480465, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01041646022349596, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17937570810317993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16898277401924133, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1587408483028412, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.15559521317481995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.14145731925964355, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07802630215883255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07446564733982086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08635454624891281, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07989876717329025, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07896017283201218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07039064913988113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06741033494472504, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.043825697153806686, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.038163721561431885, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03733640909194946, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03714500740170479, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02188876084983349, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.019462453201413155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.01940724439918995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01797129586338997, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.017939593642950058, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.011664582416415215, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012118760496377945, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.011395875364542007, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008535911329090595, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16898277401924133, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21448853611946106, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20151382684707642, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19761797785758972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17963698506355286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09911981225013733, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09460978955030441, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10954873263835907, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10141105949878693, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10031437873840332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08941053599119186, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08548305928707123, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.055555880069732666, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0482705719769001, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04724714905023575, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.047001440078020096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027707355096936226, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02416313998401165, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024097485467791557, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022201327607035637, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02216634713113308, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01448243111371994, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014351881109178066, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014145346358418465, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009290195070207119, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17963698506355286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.19387978315353394, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1787276715040207, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.17357447743415833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.15580882132053375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.08715783059597015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.082354836165905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09842923283576965, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.0906292125582695, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08861672133207321, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07785427570343018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07378388196229935, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04967241361737251, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04302673414349556, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04153674840927124, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04118208587169647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.024856215342879295, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02162039838731289, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.021502675488591194, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.019851339980959892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.019662393257021904, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01340414211153984, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013525323942303658, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01294165849685669, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009450350888073444, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1787276715040207, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.05725454166531563, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.05265740305185318, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05037189647555351, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0449763759970665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.02623250149190426, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02418726682662964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03193410858511925, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.028493454679846764, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.026702281087636948, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.023051276803016663, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02187521383166313, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.016177456825971603, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01360489334911108, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.012607249431312084, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.012360832653939724, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.008112304843962193, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00661434093490243, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.006505913101136684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.005994926206767559, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.005877996329218149, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004308421164751053, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004292026162147522, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0038680376019328833, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0030083265155553818, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.05725454166531563, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05215277150273323, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04771977290511131, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04518076404929161, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04012458026409149, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.023442059755325317, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.021347297355532646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.028156401589512825, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.026137491688132286, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.023907650262117386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02051771804690361, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.019483467563986778, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.014125102199614048, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.012409633956849575, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011231430806219578, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01093315426260233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007074777502566576, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.005827093031257391, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.005688675679266453, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005251903086900711, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005094923544675112, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0037230318412184715, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003756740363314748, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0033584353514015675, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002494887448847294, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05215277150273323, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.17641586065292358, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.16385523974895477, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1596854329109192, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1434272825717926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08138713985681534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.07676319777965546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.09188549965620041, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08426149934530258, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08260320872068405, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07213971018791199, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.06845701485872269, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04654967039823532, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04010438919067383, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.038803208619356155, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.03848539665341377, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.023208558559417725, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.019684167578816414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01957772672176361, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01778264530003071, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.017690559849143028, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.011889912188053131, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0115210497751832, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.011334659531712532, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007028370629996061, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.17641586065292358, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18076999485492706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15609408915042877, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14768363535404205, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12059248983860016, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08161270618438721, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07226625084877014, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09621713310480118, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08792365342378616, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08445827662944794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06461253762245178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05772162973880768, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04917750880122185, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.042442332953214645, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03962971270084381, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.038944557309150696, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024967888370156288, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.021438568830490112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02123129926621914, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01830521784722805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018012842163443565, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01418654527515173, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014659757725894451, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013356732204556465, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011192934587597847, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18076999485492706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15156793594360352, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1424134224653244, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.13949017226696014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.12695196270942688, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07087184488773346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06764876842498779, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.07873169332742691, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07269975543022156, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07171779870986938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06406626105308533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06145430728793144, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04055192694067955, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03562992066144943, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03482041880488396, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.034641481935977936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.020598789677023888, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.019546661525964737, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.01948535069823265, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01834205910563469, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.018306700512766838, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.011984417214989662, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013916333205997944, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.011752419173717499, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011544370092451572, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15156793594360352, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.22543247044086456, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.21190932393074036, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2077982872724533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18899723887443542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1051718071103096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10039934515953064, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11648587882518768, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10773994028568268, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10643527656793594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09495604038238525, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09095822274684906, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05958268791437149, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.051972635090351105, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05083005502820015, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05056586116552353, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.029843706637620926, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02721688523888588, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027141669765114784, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025288913398981094, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025239786133170128, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.0163539107888937, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017901424318552017, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01598644256591797, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013668502680957317, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18899723887443542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.013013681396842003, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.01288075651973486, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.005230325739830732, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.005040508694946766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.004705240484327078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.002414732240140438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.013827112503349781, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.012631618417799473, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.004727405961602926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.004606268368661404, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.00466513354331255, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.004621603526175022, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.004494037013500929, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0032803320791572332, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0021021883003413677, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.0032378616742789745, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0020036150235682726, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.0013869721442461014, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.0019871415570378304, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0013694465160369873, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.001973191974684596, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.0019767249468714, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.0013346815248951316, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.001339656999334693, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.013013681396842003, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.05935772508382797, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.05456215888261795, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05218121036887169, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04656192660331726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.027289656922221184, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02513580583035946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03208165988326073, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.029648223891854286, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.027786346152424812, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.023940660059452057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.022791458293795586, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.016239594668149948, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.014148159883916378, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.013114061206579208, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.012861385010182858, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.008139819838106632, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.006850664969533682, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00673873582854867, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006191792897880077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006069847382605076, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00430888868868351, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004408218897879124, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00397023418918252, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0030428471509367228, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.05935772508382797, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.054536182433366776, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04996170476078987, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.047497205436229706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04217386245727539, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.024753611534833908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.022630803287029266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.02942807786166668, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.027310753241181374, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.025231752544641495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02163032628595829, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.020561175420880318, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.014812290668487549, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.012989550828933716, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01186917070299387, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011595605872571468, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007419619709253311, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006147023756057024, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006020641885697842, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005528823472559452, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005388803780078888, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0039030746556818485, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003924242220818996, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003563267644494772, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0026114662177860737, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.054536182433366776, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20838594436645508, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19347314536571503, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18865670263767242, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16947461664676666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0967247486114502, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09122323989868164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10889864712953568, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09998613595962524, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09821814298629761, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08560992777347565, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08115275949239731, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05518007278442383, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04760969430208206, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.046160146594047546, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04581579566001892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02751159481704235, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02334810420870781, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023236572742462158, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021038811653852463, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.020951686426997185, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014077614061534405, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013564269989728928, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01347210444509983, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0081686582416296, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19347314536571503, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1500432789325714, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11804767698049545, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1064392477273941, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09430793672800064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0660700872540474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05191495642066002, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08094411343336105, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07458927482366562, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07034160941839218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.050814155489206314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04780363664031029, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.041531287133693695, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03602355346083641, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.032188575714826584, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.031231068074703217, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020895931869745255, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0175381638109684, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017274068668484688, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014765088446438313, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014224454760551453, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.011615113355219364, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012377044185996056, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0104636549949646, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009418363682925701, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1500432789325714, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16383898258209229, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.15391908586025238, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.15088723599910736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.13722297549247742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07641918212175369, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0728917270898819, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08461306244134903, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.0782289206981659, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07734331488609314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0688505470752716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06588756293058395, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04311663657426834, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.037525299936532974, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03671571984887123, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03652569651603699, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.021570798009634018, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.01927281729876995, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.019222870469093323, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.017802977934479713, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.017779238522052765, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.011588966473937035, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012183021754026413, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.011331231333315372, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008798599243164062, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16383898258209229, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23660315573215485, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22234201431274414, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2180994302034378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19830121099948883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11014164239168167, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10508525371551514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12168935686349869, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11269429326057434, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1114751547574997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09922043979167938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09479252249002457, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.061732787638902664, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.053679194301366806, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05253928154706955, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.052264124155044556, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03080032952129841, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.026747791096568108, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0266735702753067, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024518536403775215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024477660655975342, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01598266512155533, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01569163054227829, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015609651803970337, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009880361147224903, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12168935686349869, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2032356709241867, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1871103197336197, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18162497878074646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16305875778198242, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09154810756444931, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08639149367809296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1036161333322525, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09526260197162628, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09312183409929276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0815838873386383, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07740489393472672, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05223101004958153, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04522901773452759, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04363222420215607, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0432446263730526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02619691751897335, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022598445415496826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022474808618426323, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02067243680357933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020468274131417274, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014173033647239208, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01397924404591322, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.0136806545779109, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009556344710290432, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1871103197336197, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04891807213425636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.045150864869356155, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.042810022830963135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.03835982829332352, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.022523237392306328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02060648426413536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.027142386883497238, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.025079278275370598, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.02291138283908367, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.019867144525051117, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.01902255229651928, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.013711797073483467, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.011965716257691383, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.010848869569599628, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01057358831167221, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0068898131139576435, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.005710028577595949, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0055772592313587666, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.005193271674215794, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00505079748108983, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0036911999341100454, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0037692177575081587, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0033474506344646215, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0026345052756369114, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04891807213425636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04707751050591469, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04324079677462578, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04074747860431671, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03629692643880844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02138562686741352, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.019417904317378998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.026026416569948196, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.024159859865903854, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.021777957677841187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.018807219341397285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.017955543473362923, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.013087949715554714, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.011507830582559109, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.010284953750669956, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.009980528615415096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.006561644840985537, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00537980068475008, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00522601418197155, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.004871855955570936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0047029308043420315, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.003501919796690345, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0035395040176808834, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003138174070045352, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0023927250877022743, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04707751050591469, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.19940462708473206, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1849721074104309, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18029403686523438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16179479658603668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09255564212799072, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0871385782957077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10470610857009888, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09588483721017838, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09393390268087387, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0818166583776474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07761166244745255, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.053015995770692825, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04565170779824257, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04412786662578583, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.043758098036050797, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026437250897288322, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.022313669323921204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022186897695064545, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02009529620409012, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.019974254071712494, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013511494733393192, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.012959357351064682, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.012847506441175938, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007722103502601385, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1849721074104309, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1710282415151596, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13858217000961304, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12616890668869019, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09687469899654388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07616466283798218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0633302703499794, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09287633001804352, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08551847189664841, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07983094453811646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05447905883193016, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05021783336997032, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04746119678020477, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.041608262807130814, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03736022114753723, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.036336228251457214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02428894303739071, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02085212990641594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02052769809961319, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017100483179092407, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01656481996178627, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013813765719532967, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015166123397648335, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012535637244582176, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012065663933753967, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1710282415151596, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17663636803627014, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.16586372256278992, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.16261950135231018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1477513462305069, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08243629336357117, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07860669493675232, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09135272353887558, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08442475646734238, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08345571905374527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07419934123754501, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07096314430236816, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04648875072598457, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04039303585886955, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03950938954949379, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03930133953690529, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0232455562800169, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0205166507512331, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.020458120852708817, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01888163946568966, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.018847404047846794, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012360590510070324, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012656396254897118, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012072596698999405, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008779854513704777, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17663636803627014, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24883314967155457, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2337407022714615, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22924567759037018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20825506746768951, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11612916737794876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11072412878274918, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12841357290744781, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11883851885795593, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.117538683116436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10447611659765244, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09978030622005463, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06521062552928925, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05666303634643555, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.055440422147512436, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.055151186883449554, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.032536692917346954, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02824677899479866, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02816755324602127, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025863131508231163, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025822458788752556, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016922537237405777, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0166044682264328, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016520148143172264, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010495693422853947, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12841357290744781, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20681029558181763, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19002418220043182, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18392649292945862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1647116094827652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09306810051202774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08754169195890427, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1059831753373146, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09746582806110382, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09471577405929565, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08271553367376328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07832254469394684, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.053430840373039246, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04623844102025032, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.044336043298244476, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04388125613331795, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02681603655219078, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0229196660220623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022752489894628525, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.020915908738970757, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020666781812906265, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01443642470985651, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014162031933665276, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013828843832015991, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009578735567629337, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19002418220043182, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.05941169336438179, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.05469425767660141, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05230972170829773, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04671643674373627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0273145604878664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.025159472599625587, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03211408108472824, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.029726896435022354, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.02779337950050831, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.023996921256184578, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.022816676646471024, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.016260704025626183, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.014181360602378845, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01311185397207737, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.012850710190832615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.008141813799738884, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.006838612724095583, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.006718805991113186, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006184432655572891, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006054969970136881, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004301714710891247, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004387188237160444, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0039602299220860004, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003000673372298479, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.05941169336438179, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.054838627576828, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05029500648379326, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.047791488468647, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04247194156050682, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.024838000535964966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.022690434008836746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.029593413695693016, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.027501938864588737, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02530740760266781, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.021757308393716812, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02068636193871498, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.014873982407152653, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.013059787452220917, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011880184523761272, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011589327827095985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007439592387527227, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006119883619248867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.005981293506920338, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005506867542862892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005352976266294718, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.003891904139891267, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0038684369064867496, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0035354217980057, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0024908052291721106, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.054838627576828, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.18429358303546906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.17088736593723297, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.16637256741523743, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.14916424453258514, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08531060814857483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0802592784166336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.09643831104040146, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08859705179929733, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08664913475513458, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0753868818283081, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07140009105205536, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04881530627608299, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04216554015874863, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.040676217526197433, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04032701253890991, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02432660572230816, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.020570220425724983, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02043871022760868, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.018517661839723587, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.018400833010673523, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.012445728294551373, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.011975083500146866, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.011839814484119415, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007153350859880447, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.18429358303546906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17889004945755005, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14908915758132935, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13835902512073517, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10771994292736053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07999248802661896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06807509809732437, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09572271257638931, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08819597214460373, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0836481973528862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06033847853541374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.052661456167697906, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.049101684242486954, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04264809563755989, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03892972320318222, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03803592920303345, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024780932813882828, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02125476486980915, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020966699346899986, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017689768224954605, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01728913187980652, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013848022557795048, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014863790012896061, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012729241512715816, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011440073139965534, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17889004945755005, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16803079843521118, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.15776409208774567, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1546671837568283, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.14056679606437683, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0786026194691658, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07491723448038101, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.0869721919298172, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08051486313343048, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07956162840127945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07070878148078918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0675647035241127, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04430487006902695, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.038540054112672806, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0376911498606205, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03749186545610428, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02213425561785698, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.01959976553916931, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.019542967900633812, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.018036916851997375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01800384186208248, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.011760888621211052, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012132236734032631, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.011482803151011467, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008460672572255135, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16803079843521118, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2479194551706314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23286613821983337, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22835466265678406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20752352476119995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11577995121479034, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11038757115602493, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12795886397361755, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11854508519172668, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11722178757190704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10415327548980713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09944123774766922, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06494884192943573, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05650279298424721, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05527129024267197, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05497821792960167, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03238392993807793, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028130896389484406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028050044551491737, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.0257464200258255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025703294202685356, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016764886677265167, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01648952253162861, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016353996470570564, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01035064086318016, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12795886397361755, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20678623020648956, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1899043321609497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18422892689704895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16456864774227142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09305240213871002, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0876871719956398, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10519520938396454, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09684170037508011, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09470342844724655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08255717158317566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07790650427341461, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05303270369768143, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.045934371650218964, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04430349916219711, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0439210943877697, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.026504576206207275, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022880084812641144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022754285484552383, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02085084468126297, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020640382543206215, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014153406023979187, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014065410941839218, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.0136392368003726, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.00950822327286005, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1899043321609497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.051872801035642624, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.04781145602464676, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.04533236473798752, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04052454233169556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.023899177089333534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.021833451464772224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.02863112837076187, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0266222283244133, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.024321019649505615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.021031511947512627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.020059239119291306, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.014501294121146202, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.012711753137409687, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.011506456881761551, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.011207005940377712, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.007274491246789694, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.006051093805581331, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0059071690775454044, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.005494039971381426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.005338043440133333, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0038875252939760685, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.003995114006102085, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0035189639311283827, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0027843096759170294, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.051872801035642624, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.051520936191082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04739651083946228, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.044649265706539154, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.039794206619262695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.023457752540707588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.021263275295495987, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.028572898358106613, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.026622822508215904, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.023875072598457336, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.020641671493649483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.019703440368175507, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.014360472559928894, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.012665498070418835, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011255905963480473, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.010906173847615719, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007183410227298737, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.005849139764904976, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.005672010593116283, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005293757189065218, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005101601593196392, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0037907601799815893, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003812461392953992, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0033715576864778996, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002518164925277233, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.051520936191082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.17993271350860596, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1668064445257187, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.16230441629886627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.14548766613006592, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08328839391469955, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.07828677445650101, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.09489329159259796, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08671746402978897, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08462207764387131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07357820123434067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0699213519692421, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04805406183004379, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04125683382153511, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.03974980488419533, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0393739677965641, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02397669292986393, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02012060582637787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01998433843255043, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0181133896112442, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.017988327890634537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01226827036589384, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01176063809543848, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.011589806526899338, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007063718046993017, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.17993271350860596, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16708292067050934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12980923056602478, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11619041115045547, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09537342935800552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07389069348573685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.058810755610466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09029066562652588, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08304868638515472, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07832430303096771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05245717614889145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04958950728178024, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.046328142285346985, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04011324793100357, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03600704297423363, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.034982796758413315, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02335154265165329, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019650455564260483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01936124823987484, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015811651945114136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0152369923889637, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013054829090833664, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013873920775949955, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011829162947833538, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010613560676574707, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16708292067050934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18384034931659698, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17262506484985352, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.169158473610878, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1537005603313446, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08598966151475906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08194496482610703, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09515166282653809, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08811239898204803, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08704845607280731, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07738089561462402, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0738869160413742, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0484561026096344, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04223018139600754, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04128587990999222, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04106323793530464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02421829104423523, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02156274951994419, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0214978139847517, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01986529864370823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01982779987156391, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012891140766441822, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013479108922183514, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012581219896674156, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009546174667775631, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18384034931659698, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24049192667007446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22585222125053406, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22142870724201202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20112977921962738, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1122581884264946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10697303712368011, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12403731048107147, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1149664893746376, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11364638060331345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10095800459384918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09629979729652405, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06291568279266357, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05478496849536896, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05356403440237045, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.053279809653759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.031359054148197174, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.027244672179222107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02716297097504139, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024923456832766533, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024876857176423073, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016160082072019577, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01594233699142933, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015746286138892174, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009952317923307419, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12403731048107147, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.188278928399086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.17311899363994598, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.16683673858642578, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.14924369752407074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.08508166670799255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.07975754886865616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09828619658946991, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09028777480125427, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08660365641117096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07543756067752838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07136020064353943, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04915003105998039, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04270622879266739, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04067784547805786, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.040176186710596085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.024664506316184998, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.021211249753832817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.020971160382032394, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.019363578408956528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.01903568021953106, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013196153566241264, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013416695408523083, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.012494557537138462, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009247043170034885, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.188278928399086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06530002504587173, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06035526096820831, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05775534361600876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05177543684840202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.030073346570134163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.027789996936917305, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03561069443821907, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03273892402648926, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0305680800229311, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02658066339790821, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.025444692000746727, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018047844991087914, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01564549282193184, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014473272487521172, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014186958782374859, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009045681916177273, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007578989490866661, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007444808259606361, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006895716767758131, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006751899607479572, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0048035988584160805, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004898077808320522, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004434481263160706, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0033939320128411055, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06530002504587173, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05574115738272667, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05139373615384102, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04831412807106972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04315091669559479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.025277094915509224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02288469485938549, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03094807267189026, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.028779411688447, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.025708498433232307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.022336594760417938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.021313894540071487, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.015514818951487541, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.013709912076592445, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.012124267406761646, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011738942004740238, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007765909656882286, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006319093052297831, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006122052203863859, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00574028305709362, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005529496353119612, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0041230893693864346, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004150795750319958, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003670663107186556, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002770442981272936, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05574115738272667, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20163537561893463, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.18749916553497314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18297721445560455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16459780931472778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09374526888132095, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08847157657146454, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10604803264141083, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09680971503257751, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09513737261295319, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08309490978717804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07884969562292099, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05375337228178978, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04611583799123764, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04471375420689583, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04436596482992172, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026803715154528618, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0226127989590168, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022501587867736816, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.020406514406204224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.020317459478974342, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013712477870285511, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013125531375408173, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013068212196230888, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00789535790681839, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.18749916553497314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19124431908130646, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15550579130649567, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14228682219982147, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10839672386646271, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08532525599002838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07076659053564072, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10401998460292816, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09532018005847931, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08977018296718597, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06095299497246742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05600839853286743, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05346282199025154, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04598328098654747, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04146268591284752, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.040346644818782806, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027036594226956367, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02232888899743557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.021973418071866035, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017810478806495667, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017229072749614716, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015137383714318275, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015374152921140194, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013731954619288445, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011420461349189281, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19124431908130646, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17943648993968964, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.16847243905067444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.16515910625457764, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.15011811256408691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08389940112829208, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07998289912939072, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09275583177804947, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08591945469379425, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08493229001760483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07553873211145401, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07226298004388809, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04732542112469673, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04125284031033516, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04035412520170212, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04014356806874275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.023701190948486328, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02121102623641491, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.021151481196284294, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01958591118454933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01955169253051281, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012773675844073296, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013454359956085682, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012489786371588707, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009749718010425568, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17943648993968964, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.235514298081398, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22113299369812012, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2168436348438263, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1970752775669098, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1098153293132782, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10469206422567368, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12131022661924362, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1124107763171196, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11118249595165253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09879720956087112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09431751072406769, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.061549268662929535, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05357526242733002, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05240514874458313, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.052134621888399124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.030680056661367416, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.026676760986447334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02659584768116474, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024417156353592873, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024375101551413536, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01585542969405651, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015636801719665527, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015471470542252064, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.00982297956943512, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12131022661924362, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21055881679058075, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19468359649181366, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1893918663263321, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16987833380699158, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09496450424194336, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0899275466799736, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10678889602422714, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.0984661877155304, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09648527950048447, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08492119610309601, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08031479269266129, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.053877007216215134, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04674249514937401, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04518427699804306, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04482070729136467, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02693324163556099, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02333122491836548, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023213276639580727, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021394766867160797, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021201930940151215, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01434070710092783, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014317093417048454, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013842684216797352, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009687907062470913, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1893918663263321, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07746484875679016, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07169429957866669, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06897240877151489, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06190729886293411, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.035703033208847046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.033194392919540405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.041407547891139984, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03827840834856033, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03628021478652954, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03157472610473633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030086562037467957, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0209594015032053, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018272604793310165, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017134038731455803, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016856079921126366, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010484091006219387, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008921196684241295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008799431845545769, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008107679896056652, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007979434914886951, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005533262621611357, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005665875505656004, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00517066428437829, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003882338060066104, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07746484875679016, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06699464470148087, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.061802323907613754, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05889645963907242, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.052668701857328415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03044017404317856, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.028006047010421753, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.036047689616680145, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03348669782280922, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03094959445297718, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.026903195306658745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025574564933776855, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018110152333974838, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01592441089451313, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014582674019038677, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014250906184315681, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00907579343765974, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007529668044298887, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007375665009021759, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006823200266808271, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006657093297690153, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004763691686093807, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004772435873746872, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004366178531199694, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0031369244679808617, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06699464470148087, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2001338005065918, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1861320286989212, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18164989352226257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16322821378707886, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0926816314458847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08756700158119202, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10460331290960312, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09574364870786667, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09403946995735168, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08222490549087524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07809959352016449, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.052953269332647324, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04557393491268158, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04419650137424469, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.043864618986845016, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026383178308606148, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.022348053753376007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022244520485401154, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.020194124430418015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.020094826817512512, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013498290441930294, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.012972632423043251, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.012893348932266235, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007799987681210041, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1861320286989212, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2044530063867569, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17921264469623566, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.17041176557540894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.14367492496967316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09292012453079224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0830807164311409, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10798194259405136, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0995291918516159, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09569031745195389, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07541511952877045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06884711980819702, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05523122102022171, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04784514755010605, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04487793892621994, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.044165100902318954, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02778308466076851, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023892490193247795, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02365620993077755, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.020560728386044502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.020260274410247803, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015288890339434147, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015846146270632744, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01438086200505495, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011668204329907894, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17921264469623566, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18080465495586395, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.16976045072078705, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.16641125082969666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.15116360783576965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08463554084300995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08066342025995255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.0936107411980629, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08671428263187408, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08568227291107178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07613253593444824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0727553591132164, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0477018840610981, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04155605658888817, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04063699021935463, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.040416136384010315, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02384541556239128, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.021197691559791565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.021136272698640823, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019523898139595985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.019487770274281502, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01269678957760334, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013220131397247314, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012397346086800098, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009331787936389446, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18080465495586395, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24718748033046722, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23215441405773163, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2276650369167328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2068549543619156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11551865190267563, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11012500524520874, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1275717318058014, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11826299875974655, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11695490777492523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10389666259288788, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09908550977706909, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0647348091006279, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.056368496268987656, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.055135566741228104, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05484233424067497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03227359429001808, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028035610914230347, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02795369178056717, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02564459852874279, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025603337213397026, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016651520505547523, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016393326222896576, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01623872108757496, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010225627571344376, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1275717318058014, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21091268956661224, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19424980878829956, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18863451480865479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1684993952512741, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09506620466709137, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.089746855199337, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10701798647642136, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09878049790859222, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09665162861347198, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.084471695125103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07957756519317627, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05386463552713394, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04680962860584259, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.045216478407382965, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04483940452337265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.026880579069256783, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02328413911163807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023159248754382133, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02123154141008854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02102542668581009, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014166134409606457, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01420423574745655, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013658493757247925, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009484093636274338, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18863451480865479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07150580734014511, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06605737656354904, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06334273517131805, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05680227279663086, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03291613608598709, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.030471986159682274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03867445886135101, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.035592712461948395, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.033461373299360275, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.029057372361421585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027747053653001785, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019554512575268745, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017000511288642883, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015816131606698036, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01553153246641159, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009803272783756256, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008260120637714863, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008129607886075974, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0075046890415251255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007366550154983997, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005193138960748911, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005293000023812056, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00481997337192297, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003647780744358897, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07150580734014511, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06201682984828949, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05711555480957031, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.053991854190826416, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0481979064643383, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02815859392285347, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.025630366057157516, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03416653349995613, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03165983036160469, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.028650179505348206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.024852752685546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.023696817457675934, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.017148436978459358, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015081594698131084, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013516637496650219, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013140905648469925, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008596515282988548, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0070550087839365005, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006864494644105434, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006401150953024626, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006192418746650219, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0045702033676207066, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004611009731888771, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004101085476577282, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00310379802249372, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06201682984828949, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23093795776367188, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2149936705827713, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21001139283180237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1892033815383911, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10770662873983383, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10177204757928848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12129153311252594, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1110282614827156, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10925902426242828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09561747312545776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09082407504320145, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06149706244468689, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05293964222073555, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05139339342713356, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05103275552392006, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03067212738096714, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026001527905464172, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02586917206645012, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023490067571401596, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023403653874993324, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015688253566622734, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015070535242557526, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015002282336354256, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009042154997587204, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1892033815383911, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18020984530448914, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14520587027072906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13341844081878662, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10450153052806854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08167281001806259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0680709034204483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09732604771852493, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.088506318628788, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0847460925579071, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05753330513834953, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05363374948501587, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05002657696604729, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04292207211256027, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03985412418842316, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0391128808259964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025496114045381546, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02177746593952179, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02154901809990406, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017430366948246956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01710737682878971, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014671583659946918, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01515431422740221, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013788118958473206, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011774119921028614, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18020984530448914, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17901477217674255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.16811582446098328, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.16479259729385376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.14972294867038727, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08386487513780594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0799243152141571, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.0927727222442627, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08590240776538849, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08490294963121414, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0754220262169838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07208513468503952, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.047311414033174515, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04117976501584053, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.040272098034620285, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0400564931333065, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.023639392107725143, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.021013502031564713, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02095576375722885, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019347352907061577, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.019312659278512, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012602683156728745, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013113324530422688, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012311327271163464, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009268008172512054, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17901477217674255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.246841162443161, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2318764626979828, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2273838222026825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2065708339214325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11543303728103638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1100359782576561, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12753364443778992, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11816372722387314, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11686387658119202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10378920286893845, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09905160963535309, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06476442515850067, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05635013431310654, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05512120947241783, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05483212321996689, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03229279816150665, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028056401759386063, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02797718718647957, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02566319890320301, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025620566681027412, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016708511859178543, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016447823494672775, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016296545043587685, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010323365218937397, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12753364443778992, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20925021171569824, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19264282286167145, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18702881038188934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16720683872699738, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09433936327695847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0890066847205162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10626892745494843, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09813536703586578, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09595569968223572, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08381778001785278, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07907150685787201, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05354360491037369, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.046508099883794785, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.044871967285871506, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04448312893509865, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.026712389662861824, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02309846505522728, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022970404475927353, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021054401993751526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020844856277108192, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014094315469264984, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014085699804127216, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013573793694376945, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.00938804354518652, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19264282286167145, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06780828535556793, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06279044598340988, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05994543433189392, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05383920669555664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03128219395875931, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.028867416083812714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.037096817046403885, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03429674729704857, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03178585693240166, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02772238850593567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0264936201274395, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018772078678011894, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.016385922208428383, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01506106648594141, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014735701493918896, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009415367618203163, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007915051653981209, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007761895656585693, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007222176529467106, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0070587522350251675, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005025836173444986, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005172806326299906, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004619661718606949, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0036242890637367964, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06780828535556793, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06519021093845367, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06022603064775467, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05676465481519699, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05084001645445824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0296300258487463, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.026994675397872925, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03610632196068764, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03355052322149277, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.030127286911010742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.026253098621964455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025111403316259384, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018102476373314857, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015918368473649025, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014212858863174915, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013792894780635834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009072616696357727, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007404306437820196, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007186830509454012, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006741252727806568, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0065056635066866875, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00482942583039403, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004839172121137381, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004329116549342871, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003226944012567401, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06519021093845367, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21822191774845123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20324160158634186, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.198433980345726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17882460355758667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10153256356716156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09590711444616318, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1147579625248909, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10489653050899506, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10298505425453186, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09024976938962936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08584615588188171, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05815940722823143, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04994796961545944, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04845042526721954, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.048074908554553986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028982073068618774, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024501975625753403, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02437721937894821, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022144541144371033, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.022053765133023262, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014803985133767128, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014204185456037521, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01411085482686758, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008524924516677856, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17882460355758667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19831882417201996, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.153618723154068, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13629648089408875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.117156982421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08840585500001907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0676666796207428, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1093016043305397, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10008944571018219, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09308136254549026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06369112432003021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06100474298000336, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05614990368485451, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.048328425735235214, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04310533031821251, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.041790079325437546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02834518998861313, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023489464074373245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.023050805553793907, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019023358821868896, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018246974796056747, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0159046221524477, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01658523455262184, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014341951347887516, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012597331777215004, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.153618723154068, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19212931394577026, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18042516708374023, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1768607646226883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16077691316604614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09016784280538559, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08592096716165543, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09983918070793152, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09238437563180923, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09127718955278397, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08111301809549332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07754231244325638, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05098235607147217, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04436594992876053, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04337962344288826, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.043142177164554596, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.025534724816679955, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022773120552301407, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0227045901119709, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020999465137720108, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.020959829911589622, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013754048384726048, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014397636987268925, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013434695079922676, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01038272026926279, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19212931394577026, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24975597858428955, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23460203409194946, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23005355894565582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2090148627758026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11686776578426361, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11137470602989197, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1291843056678772, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11966242641210556, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11831847578287125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10504938662052155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10024333745241165, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0655776783823967, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05707137659192085, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.055803172290325165, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.055504415184259415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03270827978849411, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02840385213494301, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028318066149950027, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025976959615945816, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025927288457751274, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016912419348955154, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016648191958665848, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016486462205648422, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010439898818731308, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1291843056678772, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21059641242027283, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19327066838741302, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1873529702425003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1671321988105774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09488753974437714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0893080085515976, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1071770116686821, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09892969578504562, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09659623354673386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08399622142314911, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07915449142456055, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.054041244089603424, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.046908654272556305, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04517684131860733, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04475871101021767, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02697010338306427, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02333093248307705, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02319151535630226, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021223081275820732, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020997008308768272, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01429134700447321, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01435394398868084, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013740094378590584, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009693965315818787, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19327066838741302, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0732816606760025, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06775838881731033, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06495822221040726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05823278799653053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03378625959157944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.031263142824172974, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.039772119373083115, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03665277734398842, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03435894474387169, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.029862534254789352, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.028518982231616974, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020147323608398438, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017518730834126472, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01624557375907898, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015932409092783928, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010095445439219475, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008489126339554787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00834718719124794, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007717282511293888, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0075662159360945225, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0053480942733585835, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005464223679155111, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004946507979184389, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0037706089206039906, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0732816606760025, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06759686768054962, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06225953251123428, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.059016935527324677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05262623354792595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030693896114826202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02804388478398323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.036965589970350266, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03429783508181572, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.031248193234205246, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02706550434231758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025756066665053368, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018551258370280266, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01631103828549385, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014695066958665848, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014310206286609173, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009276343509554863, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007615105714648962, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007420323323458433, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006892392411828041, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006680731661617756, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004881100729107857, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004896221682429314, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004394026007503271, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032203705050051212, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06759686768054962, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.19782711565494537, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.18378904461860657, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.17907631397247314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1609342098236084, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09163979440927505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0863211378455162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10382910072803497, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09502116590738297, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09300489723682404, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08122067153453827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07715696096420288, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05263826251029968, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04524218291044235, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0436963252723217, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04333564639091492, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026249079033732414, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02216014266014099, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022036468610167503, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.020018700510263443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.019905975088477135, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013448567129671574, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.012979326769709587, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.012776081450283527, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007919788360595703, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.18378904461860657, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1612260639667511, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13695873320102692, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12863951921463013, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10877837240695953, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07304281741380692, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06351141631603241, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08553225547075272, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07866320759057999, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07564297318458557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.056510284543037415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05376644432544708, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.043938055634498596, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03800271451473236, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0354665145277977, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03485536947846413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022228866815567017, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01916433870792389, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018980927765369415, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016098473221063614, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015807008370757103, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012505770660936832, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01308183278888464, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011772889643907547, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009975177235901356, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1612260639667511, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17168693244457245, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1612449288368225, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1580624282360077, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1436852365732193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08045549690723419, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07667462527751923, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08903145790100098, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08244423568248749, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08144815266132355, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07239216566085815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06919437646865845, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04540907219052315, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03952096775174141, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.038630273193120956, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.038426052778959274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02269899658858776, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.020156674087047577, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.020092282444238663, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.018565891310572624, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.018530145287513733, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012109692208468914, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012577328830957413, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.011824044398963451, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008883032016456127, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17168693244457245, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23713018000125885, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22284308075904846, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2185641974210739, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1987813115119934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11093880981206894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10576492547988892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12256105244159698, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11358138173818588, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11231644451618195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09983158111572266, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09527713805437088, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.062212057411670685, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.054163966327905655, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05297435447573662, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05268697068095207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03103245049715042, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02695968560874462, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02687801606953144, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024676982313394547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02463381178677082, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016035471111536026, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015801478177309036, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015632642433047295, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009909944608807564, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12256105244159698, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20148880779743195, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18594355881214142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1807209551334381, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1623470038175583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09101897478103638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08601999282836914, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10258322954177856, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09453228861093521, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09253016114234924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08118069171905518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0769246518611908, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05170781910419464, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0449085459113121, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04336918145418167, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.043011248111724854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02587144635617733, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022462470456957817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.0223428625613451, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02056512050330639, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020369919016957283, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013782699592411518, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013882477767765522, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013297384604811668, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009502658620476723, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18594355881214142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07318755984306335, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06782645732164383, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06503607332706451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05845266953110695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03375783562660217, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03129138797521591, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.039510779082775116, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03658376634120941, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.034296486526727676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.029936378821730614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.028547506779432297, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019995955750346184, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017472749575972557, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016220981255173683, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015911957249045372, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01002009492367506, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008471989072859287, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00833138357847929, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007719360291957855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007571352180093527, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005306468345224857, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00544321583583951, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004915732890367508, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0037536476738750935, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07318755984306335, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06839960813522339, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0632258877158165, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05981629714369774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05365534871816635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0311738308519125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0284647848457098, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03767656162381172, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.035041943192481995, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03168429434299469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027645407244563103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026402724906802177, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01891985721886158, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01664593629539013, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014936070889234543, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014515724033117294, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009456997737288475, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007735413499176502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007520835381001234, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007037345785647631, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006799975410103798, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005005447193980217, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004981645382940769, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00450065778568387, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032480782829225063, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06839960813522339, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2384621649980545, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22259649634361267, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21751458942890167, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19634860754013062, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11128689348697662, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1053471565246582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12551814317703247, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11471129208803177, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11285664141178131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09904354810714722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0944024994969368, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0637655034661293, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05464288964867592, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05307668447494507, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05271884426474571, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03179457411170006, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02685757912695408, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0267446618527174, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024316484108567238, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024232830852270126, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016259025782346725, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01554664596915245, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015510894358158112, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009330176748335361, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12551814317703247, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18391713500022888, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1491624116897583, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13662061095237732, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11393503844738007, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08266741037368774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0679384171962738, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09990006685256958, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09107131510972977, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0862199068069458, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06244836002588272, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05683395639061928, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05090497061610222, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04412132129073143, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04035327583551407, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03943478316068649, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02587241306900978, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022160714492201805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.021861424669623375, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018545374274253845, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018060393631458282, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01481194794178009, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015659339725971222, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013714536093175411, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012199445627629757, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18391713500022888, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17727318406105042, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.16655468940734863, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.16329960525035858, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.14850443601608276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08314654976129532, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07925376296043396, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.0919368639588356, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08515968918800354, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08417133241891861, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07482823729515076, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07155774533748627, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04690985754132271, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.040845707058906555, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03994368761777878, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.039731644093990326, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.023440450429916382, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.020866239443421364, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.020804855972528458, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019227929413318634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.019192613661289215, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012501238845288754, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013056304305791855, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012210063636302948, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009262224659323692, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17727318406105042, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24115389585494995, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22669868171215057, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22238093614578247, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20232681930065155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1129511296749115, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10769219696521759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12478513270616531, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11560215055942535, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11433275789022446, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10164817422628403, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09709756821393967, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06339099258184433, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.055136922746896744, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0539400689303875, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05366039276123047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03160793334245682, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02747008018195629, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027390120550990105, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02515367977321148, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025107502937316895, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016375843435525894, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016124172136187553, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01597817987203598, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010147664695978165, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12478513270616531, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2031533122062683, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1875508427619934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18221339583396912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16385787725448608, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09186556935310364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08678299188613892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10385990142822266, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.0955297201871872, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09339837729930878, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08197134733200073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07781089842319489, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05231671780347824, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.045390065759420395, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.043809689581394196, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04343070089817047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.026229411363601685, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022721469402313232, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022596420720219612, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.020811287686228752, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020609116181731224, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014109200797975063, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014089252799749374, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013612755574285984, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009688977152109146, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1875508427619934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07597436755895615, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07057688385248184, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0678575411438942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.061175379902124405, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0352160669863224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03275831788778305, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04108448326587677, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03794439509510994, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.035765498876571655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.031314603984355927, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.029956461861729622, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020819149911403656, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018163174390792847, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016942497342824936, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016645794734358788, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010442400351166725, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008861077949404716, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00872603990137577, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008096013218164444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007953851483762264, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005548567045480013, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005698191933333874, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005173186305910349, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003962427377700806, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07597436755895615, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07213056832551956, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06686469167470932, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06335300952196121, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05689781531691551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03297131136059761, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03021601215004921, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03970490023493767, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03684486076235771, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.033488716930150986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.029287394136190414, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02803046815097332, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019946930930018425, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017507782205939293, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015826774761080742, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01539977453649044, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009993680752813816, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008197235874831676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007989715784788132, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007466181181371212, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007242830004543066, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0052683427929878235, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005274469032883644, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00477076368406415, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0034776139073073864, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07213056832551956, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24708950519561768, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23074126243591309, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2255067229270935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20397885143756866, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11552637815475464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10932212322950363, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13030461966991425, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11909537017345428, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11712455004453659, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1029844731092453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09834624826908112, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06614252179861069, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.056750476360321045, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.055147506296634674, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05476120859384537, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.032998181879520416, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027869369834661484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027751687914133072, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.025256972759962082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02515915222465992, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016852039843797684, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016104649752378464, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016098549589514732, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00961250439286232, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13030461966991425, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18908609449863434, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13722318410873413, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11860308796167374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10397548228502274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08386963605880737, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06356283277273178, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1019880622625351, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09377700835466385, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08888841420412064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05579094961285591, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05508793145418167, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0524662584066391, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04542270675301552, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04097430408000946, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.039870064705610275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02668755315244198, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022439412772655487, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02214011922478676, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017464125528931618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.016743754968047142, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015254838392138481, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015870649367570877, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01397861447185278, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012244150973856449, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18908609449863434, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1892257034778595, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1779545396566391, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.174464151263237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.15879714488983154, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08877596259117126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08464117348194122, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09829572588205338, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09089933335781097, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08985406905412674, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0799325555562973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0765681266784668, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.050172727555036545, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0436074323952198, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.042649202048778534, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04242515191435814, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.025121035054326057, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022290481254458427, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02222711592912674, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020552368834614754, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.020515942946076393, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01349473837763071, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013955583795905113, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013190987519919872, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009919623844325542, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1892257034778595, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.242406964302063, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2279471755027771, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22360749542713165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20352576673030853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11353383213281631, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10826634615659714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12547025084495544, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11618897318840027, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1149178221821785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10223401337862015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09770694375038147, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06376522034406662, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.055435359477996826, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05422867834568024, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05394614860415459, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03180743008852005, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02763330191373825, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027551820501685143, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025309400632977486, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025269463658332825, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016500867903232574, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0162380151450634, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016103345900774002, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010255257599055767, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12547025084495544, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20541724562644958, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18994008004665375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18462584912776947, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16669948399066925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09293767809867859, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08790863305330276, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10503893345594406, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09664783626794815, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09448467195034027, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08326119929552078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07925919443368912, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0529380738735199, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0458756685256958, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04427400603890419, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04389381408691406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02652151510119438, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022852245718240738, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022725386545062065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.020974131301045418, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020769067108631134, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014237911440432072, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014024587348103523, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013737194240093231, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009459794498980045, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18994008004665375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07399505376815796, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06887812167406082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06605087220668793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.059613119810819626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03431900590658188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03187216445803642, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04026497155427933, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.037312138825654984, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03483867645263672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030593089759349823, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.029243772849440575, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020372331142425537, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01783645898103714, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01652826927602291, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016206104308366776, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01024662982672453, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00867606420069933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00852489285171032, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007943949662148952, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007787622511386871, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005509841721504927, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005631405860185623, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005122445523738861, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003942994400858879, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07399505376815796, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.072444386780262, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06726352870464325, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06395544856786728, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05751373618841171, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.033223673701286316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03055768832564354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.039809148758649826, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03694623336195946, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03374026343226433, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.029577022418379784, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.028338095173239708, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020013172179460526, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017590729519724846, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015947984531521797, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01553709339350462, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010037969797849655, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008251221850514412, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008055276237428188, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0075204442255198956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007310085464268923, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005310450214892626, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00527610257267952, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004845170769840479, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0034726918675005436, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.072444386780262, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23455478250980377, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21937018632888794, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21410447359085083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19389061629772186, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10966421663761139, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10379017144441605, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1240532249212265, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11347154527902603, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1111849844455719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0978870540857315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09344396740198135, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06290173530578613, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05412270128726959, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.052404507994651794, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05200409144163132, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.031352683901786804, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026576660573482513, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026434123516082764, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024133583530783653, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024015871807932854, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016066033393144608, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01553436554968357, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015319358557462692, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0094935093075037, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1240532249212265, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18751393258571625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13887372612953186, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12218397855758667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10490009188652039, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08333480358123779, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06328621506690979, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10120884329080582, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0920826867222786, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08793248981237411, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05710270255804062, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05496738851070404, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05242839455604553, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04488017410039902, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04107685014605522, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04010177403688431, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027988983318209648, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02296658605337143, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02272966504096985, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0184146910905838, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017887236550450325, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01776261441409588, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016617702320218086, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.016863226890563965, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013349970802664757, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18751393258571625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19982559978961945, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18780839443206787, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18418002128601074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1676689237356186, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0937916710972786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08943821489810944, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10383576899766922, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09602239727973938, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09493957459926605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08447622507810593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08089053630828857, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05300189554691315, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04607156291604042, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04506939277052879, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0448358952999115, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.026497874408960342, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02355441451072693, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.023488424718379974, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.021716265007853508, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.021677864715456963, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014187091961503029, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014737256802618504, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013866453431546688, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010475738905370235, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18780839443206787, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24959881603717804, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23474851250648499, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23026317358016968, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20961393415927887, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11698026955127716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11156325787305832, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12934954464435577, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.119716577231884, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11842673271894455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10535392165184021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10073485970497131, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06578997522592545, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.057131603360176086, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05589720979332924, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05560733750462532, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03283030167222023, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028520755469799042, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0284385085105896, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02613503858447075, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02609323337674141, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01711612194776535, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016823166981339455, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016707008704543114, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010710879229009151, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12934954464435577, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21094010770320892, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19443023204803467, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18874187767505646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17055216431617737, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09540951251983643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08998264372348785, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10797398537397385, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09942229092121124, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09707296639680862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08526752144098282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08130166679620743, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05443781986832619, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.047192808240652084, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04546466842293739, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.045051492750644684, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027233518660068512, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.023501351475715637, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.0233616903424263, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021531403064727783, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021305281668901443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014561129733920097, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014477452263236046, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014012337662279606, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009809473529458046, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18874187767505646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06575977802276611, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06108790263533592, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05817319452762604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05240507051348686, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.030408624559640884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02802140824496746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03627927973866463, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03363627940416336, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.030890539288520813, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02706468291580677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.025927633047103882, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018348328769207, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.016082046553492546, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01465180516242981, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014296021312475204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009200239554047585, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007700832560658455, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007531058043241501, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0070489877834916115, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006867802236229181, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004915422294288874, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0050629205070436, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0044854977168142796, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0035361917689442635, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06575977802276611, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06441130489110947, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05968542397022247, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05594012513756752, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.050196487456560135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.029373997822403908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0265981275588274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.036393534392118454, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03379948437213898, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.029854100197553635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.026137588545680046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02506171725690365, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01818433403968811, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016088826581835747, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014123550616204739, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013628841377794743, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009133460931479931, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0073699443601071835, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007114256266504526, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006728965789079666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006450026761740446, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004846720024943352, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004879736807197332, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004278768785297871, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00325208087451756, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06441130489110947, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23382259905338287, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2185635268688202, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21346260607242584, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19303040206432343, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10931470990180969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10338824987411499, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12376350909471512, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11291860789060593, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11077115684747696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09742677211761475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09293271601200104, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06282100081443787, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.053849369287490845, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05220097303390503, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.051811445504426956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03130873292684555, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026428455486893654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026299629360437393, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02396847866475582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023856883868575096, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016014546155929565, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015372288413345814, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015226228162646294, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00927063636481762, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19303040206432343, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.20590896904468536, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16412943601608276, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15016306936740875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1054748147726059, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09221400320529938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07642050087451935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10997337847948074, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10081421583890915, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09670178592205048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06110750511288643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.056362029165029526, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0567120723426342, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04868875443935394, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04480152204632759, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04384898766875267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02866113930940628, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.024193838238716125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.023938816040754318, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018569061532616615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018218649551272392, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.016114959493279457, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016601119190454483, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014971772208809853, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012537984177470207, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16412943601608276, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20735277235507965, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19500437378883362, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19123300909996033, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17401687800884247, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09727910161018372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09278121590614319, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10769955813884735, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09960155189037323, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09846797585487366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08761964738368988, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08386220037937164, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05492047592997551, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.047730203717947006, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04669566825032234, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.046451713889837265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027457810938358307, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02431376837193966, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024246307089924812, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022401917725801468, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02236393466591835, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014656265266239643, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015091457404196262, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014324394054710865, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010582081973552704, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19123300909996033, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.251859188079834, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2369103878736496, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2323811948299408, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21156218647956848, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11802957952022552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11255436390638351, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13047967851161957, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1207614615559578, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11945286393165588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10629693418741226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10164892673492432, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06635864078998566, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05763671547174454, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05639444291591644, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.056104108691215515, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03312025964260101, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02875915914773941, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02867783233523369, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.0263536274433136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026311276480555534, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01725981943309307, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016936656087636948, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016853492707014084, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01075681671500206, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13047967851161957, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2174353003501892, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2000522017478943, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19402241706848145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.175445556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0982060357928276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09250824898481369, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1110573261976242, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10242269188165665, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09998167306184769, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08768957853317261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08362697809934616, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0559200718998909, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04856215417385101, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0467509850859642, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.046321265399456024, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027955004945397377, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024069027975201607, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023924000561237335, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022019827738404274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021779509261250496, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014839230105280876, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014702248387038708, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014272057451307774, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009783649817109108, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.175445556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.062279973179101944, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.057955630123615265, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.055039454251527786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0496930330991745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.02885703183710575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.026536865159869194, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03456917032599449, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.032117247581481934, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.029292946681380272, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.025737572461366653, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.024677783250808716, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.017465809360146523, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.015374084003269672, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0139269158244133, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.013569512404501438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00877536740154028, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00736579205840826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007191926706582308, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006762710399925709, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006580257322639227, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004722051788121462, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00490986555814743, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00430668517947197, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0034904603380709887, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.062279973179101944, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0631251335144043, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.058635517954826355, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.055004432797431946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04951227828860283, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.028926456347107887, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.026244906708598137, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0357743576169014, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03325507044792175, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02938639372587204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02578100375831127, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02480347268283367, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.017958762124180794, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01581435091793537, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013922026380896568, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013446683064103127, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009019304066896439, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007279481738805771, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007032666355371475, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006659657694399357, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006396482232958078, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004820408299565315, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004824426025152206, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004286815412342548, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032433466985821724, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0631251335144043, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23843131959438324, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2227248251438141, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21754032373428345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19672717154026031, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11143718659877777, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10551055520772934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1260397583246231, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11519117653369904, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1130463108420372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09954164922237396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09496897459030151, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06404750049114227, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.054997581988573074, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05324072763323784, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.052838414907455444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03193674236536026, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027007009834051132, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026865659281611443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02451499179005623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024393808096647263, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016368716955184937, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01576947048306465, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015581714920699596, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009603112936019897, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1260397583246231, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19540195167064667, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14398786425590515, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1249297633767128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10534973442554474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08581800013780594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06322434544563293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1066713035106659, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09781528264284134, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09195229411125183, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.059776753187179565, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.057512715458869934, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05485927313566208, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04720863327383995, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04180770739912987, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04046490043401718, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027857953682541847, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022699862718582153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02232000231742859, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0180024616420269, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0172140970826149, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01573343016207218, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01598912477493286, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014129896648228168, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012019624002277851, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14398786425590515, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21359951794147491, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20090584456920624, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19702735543251038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.179338276386261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10026677697896957, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09563221782445908, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1108984649181366, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10264766961336136, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10148905217647552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09032391756772995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08638641238212585, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05654940381646156, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04918365553021431, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.048128366470336914, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.047879576683044434, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.028264444321393967, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02504684403538704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024978043511509895, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.023070737719535828, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02303621731698513, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015049304813146591, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01552534382790327, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014706957153975964, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010865814983844757, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.179338276386261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2565324902534485, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24127374589443207, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2367144674062729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21545599400997162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1202169731259346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11467120051383972, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13282807171344757, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12300229072570801, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12167730927467346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10827989131212234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10346551984548569, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06751827150583267, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.058694351464509964, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05743725970387459, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05713842436671257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.033683449029922485, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029276270419359207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029196780174970627, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02682529017329216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02678242139518261, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017502296715974808, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0172213613986969, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017088042572140694, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010903975926339626, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13282807171344757, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2205260843038559, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20258522033691406, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19627639651298523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1774950921535492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09946704655885696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0935727059841156, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11247837543487549, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10390758514404297, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10130465030670166, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08873657882213593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08463553339242935, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05655501410365105, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.049198441207408905, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04729609191417694, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.046846095472574234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02820037491619587, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024252086877822876, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024101894348859787, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.0221533365547657, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02190231904387474, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014780318364501, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01468765176832676, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014176337979733944, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009587760083377361, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1774950921535492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06941172480583191, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06453879177570343, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.061532579362392426, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0554250031709671, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03215596452355385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.029655277729034424, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.038264382630586624, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0355091318488121, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03264368698000908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02860932983458042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02737279236316681, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019348205998539925, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.016978712752461433, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0155071672052145, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015143327414989471, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009749063290655613, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00818516593426466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008009425364434719, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007500078994780779, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007313236594200134, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005279106553643942, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005413493141531944, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0048509868793189526, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003834092989563942, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06941172480583191, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0692545473575592, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0642126277089119, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06053056940436363, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05438948795199394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03162675350904465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.028847569599747658, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03860223665833473, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03587212413549423, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03212565556168556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.028123147785663605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026990259066224098, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01932891458272934, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017022093757987022, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015170089900493622, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014717013575136662, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009681529365479946, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007869149558246136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007644285447895527, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007175569888204336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00692879781126976, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0051249247044324875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00510827312245965, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00460234098136425, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0033666694071143866, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0692545473575592, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2367977648973465, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22095827758312225, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2157689779996872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1950850784778595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1103869378566742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10450678318738937, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1249007061123848, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1139826774597168, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11200354993343353, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0984477624297142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09389549493789673, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06336337327957153, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05432189628481865, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05269695073366165, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0523303747177124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03158850595355034, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02667541801929474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026562390848994255, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0241951122879982, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02408553846180439, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016151703894138336, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015510636381804943, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015393772162497044, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009373783133924007, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1249007061123848, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2007996290922165, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15997134149074554, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14554333686828613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11264336854219437, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08912747353315353, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07275383174419403, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10801186412572861, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09930911660194397, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.094444639980793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06316717714071274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.058496758341789246, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05536544695496559, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04799951612949371, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04338669404387474, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.042258355766534805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02798772044479847, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023577259853482246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02326405793428421, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018875906243920326, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018263917416334152, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01566486991941929, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016441088169813156, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014311556704342365, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012468460947275162, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15997134149074554, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2164326161146164, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20337702333927155, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19945290684700012, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18144799768924713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10162964463233948, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09690090268850327, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11232225596904755, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10402471572160721, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10285825282335281, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0914989486336708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08749173581600189, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.057283055037260056, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04983975738286972, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04876358434557915, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04850279539823532, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02862461842596531, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02534276805818081, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025270512327551842, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.023329632356762886, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023289131000638008, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01520371250808239, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01566273160278797, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014853780157864094, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010907758958637714, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18144799768924713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2587037682533264, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24326325953006744, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23862417042255402, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21712440252304077, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12126605957746506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11563051491975784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13392530381679535, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12408535182476044, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12274155765771866, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10917157679796219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10428585112094879, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06804117560386658, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05919953063130379, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05792680382728577, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.057625532150268555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.033964429050683975, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029502909630537033, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0294188242405653, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02701648138463497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02697155997157097, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017626045271754265, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01731434278190136, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017203230410814285, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010905989445745945, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13392530381679535, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22107838094234467, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20230668783187866, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1957707405090332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17681193351745605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09963875263929367, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09348898380994797, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11286551505327225, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10420945286750793, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10156292468309402, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08857355266809464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08447161316871643, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05689495429396629, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04938130080699921, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0474090538918972, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04694308713078499, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02840774692595005, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024396082386374474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02423986606299877, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022241177037358284, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02198229543864727, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015040915459394455, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014907905831933022, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014418499544262886, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009886114858090878, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17681193351745605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06281562149524689, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.05835743248462677, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05542374774813652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.049890853464603424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.028996333479881287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02665923349559307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03473704680800438, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0322844497859478, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.02943992055952549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02578853629529476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02468150667846203, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.017509883269667625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.015375321730971336, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.013940242119133472, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01358700729906559, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00877299439162016, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007293013855814934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007117183413356543, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006664672400802374, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006478258408606052, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004665771499276161, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0047599151730537415, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004240596201270819, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003255876014009118, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06281562149524689, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06639764457941055, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06157305836677551, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05833427235484123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05243248492479324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03039608709514141, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.027836579829454422, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03670266270637512, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03407377377152443, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.030862972140312195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0270290095359087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025886494666337967, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018446918576955795, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016214407980442047, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014584549702703953, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014175442978739738, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009256393648684025, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007554715499281883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007353958673775196, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0068874116986989975, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006670818664133549, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0048831757158041, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004865903872996569, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004402379970997572, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0031939190812408924, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06639764457941055, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20656916499137878, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1927236169576645, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1878696084022522, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16963157057762146, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09594417363405228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09062493592500687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10885249823331833, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09960286319255829, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09730225056409836, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08553536236286163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08157946914434433, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.055125683546066284, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04742927849292755, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.045781783759593964, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04538125917315483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027457062155008316, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023189937695860863, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023044919595122337, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021038876846432686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.020915767177939415, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014066837728023529, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013551507145166397, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013361593708395958, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008196424692869186, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1927236169576645, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18730445206165314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14996908605098724, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13605839014053345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10965096950531006, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08347049355506897, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06760524213314056, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10210800915956497, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09331955015659332, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08763846755027771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06103084236383438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05600413307547569, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.052199795842170715, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.045052189379930496, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04066212475299835, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.039572544395923615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02637510374188423, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022161398082971573, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.021809261292219162, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018102116882801056, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017566144466400146, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014825661666691303, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015558579936623573, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013497554697096348, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011887745000422001, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18730445206165314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21551525592803955, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20250563323497772, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19851243495941162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18057970702648163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1010713130235672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09633921086788177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11183971911668777, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10354665666818619, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10231739282608032, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09096448123455048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08697233349084854, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05698525160551071, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04961076006293297, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04851175844669342, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04825204238295555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.028488993644714355, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02526055835187435, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025185128673911095, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02325533889234066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023209309205412865, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015147029422223568, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015677999705076218, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014787809923291206, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01097836159169674, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18057970702648163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25397777557373047, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23864778876304626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23403948545455933, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21285827457904816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11888366937637329, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1132928878068924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13135841488838196, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12171900272369385, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12034167349338531, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10695356130599976, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10210838913917542, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06671507656574249, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05804573744535446, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05675949528813362, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05645552650094032, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03325987234711647, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028884852305054665, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028799299150705338, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026432082056999207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026384463533759117, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017176326364278793, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016916433349251747, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016742633655667305, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010593879967927933, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13135841488838196, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2123328596353531, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19406954944133759, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18762023746967316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16898804903030396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09572291374206543, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08971505612134933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10872066020965576, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10031361877918243, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09759137034416199, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0849609300494194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08082783222198486, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05488549917936325, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.047668300569057465, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04569024592638016, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04521828517317772, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027461063116788864, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.023820342496037483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023662041872739792, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021770158782601357, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02151232399046421, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014756957069039345, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015002468600869179, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014142429456114769, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010489225387573242, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18762023746967316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07301857322454453, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06785128265619278, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06506683677434921, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05861508846282959, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03380196914076805, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0313701406121254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03974718973040581, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03668127954006195, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.034318678081035614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03007573075592518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.028806284070014954, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02014455758035183, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01755760982632637, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016277285292744637, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015963103622198105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010117881000041962, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008535433560609818, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008391180075705051, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0078095076605677605, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007656551897525787, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005395934917032719, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005528355948626995, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005013022106140852, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0038620491977781057, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07301857322454453, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07025966048240662, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06507209688425064, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06177198886871338, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.055336710065603256, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03198989853262901, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.029331251978874207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03866558149456978, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03572451323270798, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03248772770166397, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.028388280421495438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.027179723605513573, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019470615312457085, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017024165019392967, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015354810282588005, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014946294948458672, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009797142818570137, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007981966249644756, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0077858916483819485, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007276727817952633, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007063779514282942, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005280138459056616, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005175153724849224, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004782863426953554, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0034677432850003242, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07025966048240662, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2163366824388504, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20179396867752075, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19703970849514008, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17804992198944092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10065438598394394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09522596001625061, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11360061913728714, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10399036109447479, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10207211971282959, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08977626264095306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08563094586133957, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.057553261518478394, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04954707622528076, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04801713675260544, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.047681860625743866, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02870633825659752, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024283435195684433, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024166665971279144, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022016946226358414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02192329801619053, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014675653539597988, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014080214314162731, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01401034090667963, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008432187139987946, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17804992198944092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1778518259525299, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13897022604942322, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12496829032897949, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09663201868534088, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0791427344083786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06279529631137848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09690149128437042, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08835428208112717, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08361469209194183, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.053687550127506256, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05163950100541115, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.049784138798713684, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04277471825480461, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03861606866121292, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03759513050317764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025382880121469498, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.021034132689237595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020737407729029655, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016468806192278862, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015884334221482277, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014372522011399269, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014758829958736897, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013116835616528988, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011262837797403336, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1778518259525299, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19151167571544647, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17998898029327393, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17651169002056122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16053900122642517, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08970215171575546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08554977178573608, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09914782643318176, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09188851714134216, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09079067409038544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08081220090389252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07724524289369583, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.050539709627628326, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04405619204044342, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0430908277630806, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.042858801782131195, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.025245027616620064, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022522222250699997, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02245592325925827, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020774750038981438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.020733332261443138, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01343668345361948, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014106745831668377, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013121734373271465, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010029911063611507, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19151167571544647, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2347722053527832, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22071319818496704, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.21649640798568726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19698862731456757, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10968347638845444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10460911691188812, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12108150869607925, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1122806966304779, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11102568358182907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09879620373249054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0943458154797554, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06142602860927582, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05351162329316139, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05234227329492569, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05206699296832085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.030614305287599564, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.026614772155880928, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.026538796722888947, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024375297129154205, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024333901703357697, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015769299119710922, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015556249767541885, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015380690805613995, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009697592817246914, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12108150869607925, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21332645416259766, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1960761547088623, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19014830887317657, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17145368456840515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09626981616020203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0906652957201004, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1089162826538086, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10041109472513199, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09801232814788818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08581547439098358, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08167119324207306, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.054956357926130295, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04764501750469208, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.045868102461099625, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04544323682785034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02746998518705368, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02375093847513199, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023610947653651237, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021739482879638672, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021507972851395607, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014670206233859062, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014703374356031418, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014111483469605446, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010036366991698742, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19014830887317657, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08033019304275513, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0745411068201065, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07177383452653885, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0646694079041481, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.037140142172575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03459687903523445, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04302602261304855, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03980550169944763, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.037727463990449905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.032975830137729645, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.031485337764024734, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02178194373846054, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01901603862643242, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017843659967184067, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01756017468869686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010906410403549671, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009309274144470692, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0091829439625144, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008489460684359074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008358555845916271, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0057808770798146725, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005934072658419609, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005420917645096779, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004101264290511608, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08033019304275513, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07737652212381363, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07166849076747894, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06836845725774765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.061403460800647736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.035335786640644073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03259150683879852, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04182322695851326, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.038830749690532684, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03587156534194946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.031350865960121155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.029918817803263664, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02100246772170067, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01846667192876339, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016917886212468147, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016535243019461632, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010513588786125183, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00872833002358675, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008543754927814007, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007939820177853107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0077425045892596245, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005521528888493776, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0055201356299221516, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005066687241196632, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0036054060328751802, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07737652212381363, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2183348536491394, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20357297360897064, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19885504245758057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.179656982421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10123887658119202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09585839509963989, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11375553905963898, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10449092090129852, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10267115384340286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09029162675142288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08600914478302002, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.057566579431295395, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.049730949103832245, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.048259902745485306, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04792111739516258, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02871716395020485, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02439090609550476, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024271175265312195, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022117413580417633, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.022020617499947548, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014681207947432995, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014112072065472603, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014061622321605682, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008425968699157238, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.179656982421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19626955687999725, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1566900908946991, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1432739794254303, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11710148304700851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08765415847301483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0714268684387207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10461771488189697, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09632986038923264, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09193532168865204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06442548334598541, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05914715677499771, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05349540710449219, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04624433442950249, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04233045503497124, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0413752943277359, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026904921978712082, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022447245195508003, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0221652053296566, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018050365149974823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017557714134454727, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014750410802662373, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01495450735092163, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013557971455156803, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010744921863079071, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1566900908946991, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20643457770347595, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19399040937423706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19021107256412506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17298831045627594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09675069898366928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09224621951580048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10690825432538986, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09905311465263367, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09793616831302643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08711057156324387, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08327408134937286, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0544753223657608, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04747770354151726, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.046451449394226074, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.046208493411540985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02720436081290245, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024219749495387077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02415512502193451, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02231985330581665, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022282585501670837, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014446972869336605, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015089103952050209, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0141197694465518, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010642309673130512, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19021107256412506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24965158104896545, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23473100364208221, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23024694621562958, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20947441458702087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11676359921693802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11136811226606369, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12884970009326935, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11945608258247375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11817450821399689, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10513125360012054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10040808469057083, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06539900600910187, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05696495622396469, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05574385076761246, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05545172840356827, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03260009363293648, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028361277654767036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028283432126045227, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025976436212658882, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025931933894753456, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016822926700115204, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016606036573648453, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016417190432548523, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010400979779660702, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12884970009326935, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2244510054588318, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20518098771572113, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19847477972507477, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1791137009859085, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10102906823158264, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09472360461950302, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11482323706150055, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10573434084653854, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10302074253559113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08975648880004883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08550119400024414, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05773443728685379, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.050109781324863434, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.048087604343891144, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.047602009028196335, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02884645387530327, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024792108684778214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024627236649394035, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022607319056987762, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02233375608921051, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015266572125256062, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015218120068311691, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01460936851799488, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010170000605285168, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1791137009859085, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06591320782899857, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.061226412653923035, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05830296874046326, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05251151695847511, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.030512074008584023, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.028124436736106873, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.036376144737005234, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03376499190926552, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.030981235206127167, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.027167711406946182, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02599945105612278, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018402626737952232, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01617560163140297, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014756053686141968, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014403091743588448, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00931338220834732, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007854136638343334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007685224059969187, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007211565505713224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007033008616417646, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005122198257595301, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0052749523892998695, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004721928853541613, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0038245657924562693, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06591320782899857, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06668153405189514, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06184947490692139, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05848421901464462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.052552059292793274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030607379972934723, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.027969492599368095, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03715530037879944, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03443455696105957, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03108352981507778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027219198644161224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026105085387825966, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018671955913305283, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016423309221863747, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014685318805277348, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014259479939937592, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009373190812766552, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007618953939527273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007402763701975346, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0069538988173007965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006715727970004082, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004974544048309326, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004941939376294613, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004475207068026066, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032612960785627365, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06668153405189514, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24387063086032867, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22764934599399567, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22243301570415497, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2010369598865509, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11396501958370209, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1078135222196579, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12851791083812714, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11752647906541824, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11550377309322357, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10152298212051392, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09680448472499847, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06512153893709183, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05600797384977341, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05440349876880646, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05401499941945076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03247872740030289, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02749311551451683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027368102222681046, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024908224120736122, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024817854166030884, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016596795991063118, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01590452343225479, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015863491222262383, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009507285431027412, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12851791083812714, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1733311414718628, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13890865445137024, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12596331536769867, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09552891552448273, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07547471672296524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0626399889588356, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09609533101320267, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08671296387910843, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08120106905698776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.054634831845760345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.051186997443437576, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.048738181591033936, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04216637462377548, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.037243377417325974, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.035987116396427155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025090109556913376, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02100232243537903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020682960748672485, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017475567758083344, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01683655194938183, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014617868699133396, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015588457696139812, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013198046013712883, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01249843742698431, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1733311414718628, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2111542969942093, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19837722182273865, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1944853961467743, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17674393951892853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09898751974105835, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09435629099607468, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10943130403757095, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10136115550994873, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10021841526031494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08906013518571854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08509878069162369, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0557435043156147, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04855140671133995, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04749225825071335, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04724443331360817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02784767560660839, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024686872959136963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024617502465844154, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02271578647196293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02267426811158657, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014757564291357994, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015261256136000156, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014412627555429935, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010626566596329212, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17674393951892853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25464069843292236, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23930206894874573, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2347138524055481, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2133733630180359, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11918053030967712, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11362006515264511, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13157829642295837, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12195146828889847, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12061377614736557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10723225027322769, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1023598164319992, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06684084981679916, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.058161575347185135, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05690615996718407, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0566120445728302, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03333346173167229, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02898850105702877, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028906753286719322, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02654007449746132, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026492711156606674, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017265617847442627, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01702186092734337, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016848178580403328, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010735217481851578, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13157829642295837, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22861360013484955, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20871393382549286, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.201717346906662, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18219023942947388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10284227132797241, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09629154205322266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11696960777044296, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10784444957971573, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10493241995573044, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09134570509195328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08708972483873367, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05889506638050079, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05108137056231499, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04892344027757645, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04840566962957382, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02944701351225376, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025154052302241325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024982374161481857, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02291908860206604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02263040654361248, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01565978117287159, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015365558676421642, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01499190367758274, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01014268584549427, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18219023942947388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07805435359477997, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07256965339183807, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0697963684797287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06289747357368469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03617309033870697, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03369259089231491, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04203132912516594, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03893957659602165, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03671197593212128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03218401223421097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030730970203876495, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021246861666440964, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018601248040795326, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01738855056464672, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01709751971065998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010651985183358192, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009092297405004501, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008960122242569923, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008310401812195778, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008173434063792229, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005672404542565346, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005834329407662153, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005313048604875803, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004057491198182106, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07805435359477997, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07718715816736221, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07167983800172806, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06827899813652039, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.061415158212184906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.035384707152843475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.032691460102796555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04215080663561821, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.039084915071725845, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03593447059392929, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03151461109519005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03016859106719494, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.021207323297858238, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01859104260802269, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.017006337642669678, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01660987176001072, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010636451654136181, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00883017759770155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008629992604255676, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008060263469815254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00784977525472641, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005637102760374546, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005659065674990416, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0051614767871797085, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003769234288483858, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07718715816736221, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24434851109981537, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2281263917684555, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22287708520889282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20137251913547516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11385174095630646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10781227797269821, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1280381977558136, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11732113361358643, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11547831445932388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10151176899671555, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09684823453426361, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06482492387294769, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05588017776608467, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.054304808378219604, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05393340438604355, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03233727812767029, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027441952377557755, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02732219360768795, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024868104606866837, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02478654496371746, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016511382535099983, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015829134732484818, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015826905146241188, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009410635568201542, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1280381977558136, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21615351736545563, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16261087357997894, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14337243139743805, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11515885591506958, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0964147225022316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07557787001132965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11640436947345734, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10713082551956177, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10123307257890701, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06350579857826233, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.061121195554733276, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.059437066316604614, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05130899325013161, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04647160321474075, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04527362436056137, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.029799634590744972, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02451704628765583, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.024131715297698975, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018330182880163193, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017550192773342133, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.016121599823236465, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01627921499311924, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014609185978770256, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011461492627859116, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16261087357997894, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21591177582740784, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20278598368167877, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19882555305957794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18067513406276703, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10140334069728851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09664562344551086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11208632588386536, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10382436215877533, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10265067219734192, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09118520468473434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08712494373321533, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.057240985333919525, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04983971267938614, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.048754896968603134, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.048497218638658524, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0286038089543581, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02553209289908409, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025461921468377113, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02353089675307274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023491594940423965, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015310468152165413, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016058169305324554, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014967916533350945, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011498389765620232, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18067513406276703, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25878795981407166, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24315804243087769, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23847566545009613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21679602563381195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12117264419794083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11549308151006699, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13381841778755188, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12399256974458694, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1226528063416481, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10895644873380661, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10399547219276428, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06793161481618881, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.059148870408535004, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05786653980612755, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05756102874875069, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03388075903058052, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029437869787216187, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02935280092060566, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026928070932626724, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026882704347372055, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017506621778011322, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017226442694664, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017080038785934448, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010770895518362522, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13381841778755188, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2313738465309143, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21141697466373444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20427671074867249, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18491430580615997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.1041533499956131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09755207598209381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11883170157670975, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10935159027576447, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10626967996358871, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09275826066732407, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08852508664131165, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05971686914563179, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.051812008023262024, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04959728941321373, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04906266927719116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029837487265467644, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02557317353785038, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025389201939105988, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02335928939282894, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023059459403157234, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015782015398144722, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01572265662252903, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015050970017910004, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010499753057956696, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18491430580615997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08147525787353516, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07578791677951813, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07292978465557098, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06576097011566162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03776700049638748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.035214848816394806, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04373560845851898, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04057268053293228, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.038338467478752136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.033627502620220184, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03208717331290245, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.022118257358670235, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.019396690651774406, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01817646250128746, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01788136549293995, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011089805513620377, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00953649077564478, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00940153282135725, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008727265521883965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008587894961237907, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00591411255300045, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0061482698656618595, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0055483547039330006, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0043273381888866425, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08147525787353516, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08035501837730408, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07459309697151184, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07124391943216324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06405225396156311, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03685136139392853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03410687670111656, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04358747974038124, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04044991359114647, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03740561380982399, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03281325474381447, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03139632195234299, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02194414660334587, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019233379513025284, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01767362654209137, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01728772185742855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011009863577783108, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009131967090070248, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008948475122451782, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00833053607493639, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008133983239531517, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005803011357784271, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005787340458482504, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005337278824299574, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0038042955566197634, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08035501837730408, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24954037368297577, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23326517641544342, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22816413640975952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20647627115249634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11661476641893387, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.110601507127285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13073565065860748, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11992087215185165, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11815045773983002, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10417424887418747, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09933951497077942, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06620636582374573, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.057173509150743484, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05563493072986603, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05526813864707947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03304078057408333, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028111424297094345, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02799486741423607, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02553888037800789, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025454405695199966, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016878001391887665, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016229188069701195, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01620878092944622, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009673051536083221, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13073565065860748, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2253950834274292, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17121626436710358, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14977414906024933, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11550243943929672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09915584325790405, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0765293687582016, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12529000639915466, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11409712582826614, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10589594393968582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06764845550060272, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0638902485370636, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06362071633338928, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05532471835613251, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.048622190952301025, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.046922050416469574, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.032384198158979416, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.026859842240810394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.026339055970311165, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021344847977161407, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.020377516746520996, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.018246842548251152, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0194492619484663, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01624261774122715, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.015051089227199554, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17121626436710358, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.22030478715896606, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20681288838386536, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2027367204427719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18420955538749695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10339910537004471, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09852025657892227, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11439665406942368, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10589183866977692, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1046728864312172, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09291902184486389, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08875887840986252, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.058318592607975006, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.050735071301460266, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04962274432182312, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04936191439628601, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.029159514233469963, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.025824423879384995, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025753235444426537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.023748155683279037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023707563057541847, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015539510175585747, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016007496044039726, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015182278119027615, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011191010475158691, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18420955538749695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2631781995296478, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24723197519779205, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.242437481880188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2202935814857483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12331923097372055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11752716451883316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13624370098114014, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12620720267295837, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12483420222997665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11085617542266846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10580195486545563, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06922587752342224, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.060226209461688995, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.058916617184877396, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05860402435064316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03455086797475815, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030026162043213844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029939061030745506, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027468569576740265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027422189712524414, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017931746318936348, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017660390585660934, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017509158700704575, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011166687123477459, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13624370098114014, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23270025849342346, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21247340738773346, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20519649982452393, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18597908318042755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10475009679794312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09804427623748779, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11916562914848328, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11004821956157684, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10692042112350464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09328477829694748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08906882256269455, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.059988729655742645, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05212264135479927, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.049835655838251114, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04929148405790329, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029912106692790985, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025644734501838684, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025458551943302155, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023424014449119568, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023113390430808067, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.0157308392226696, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015708012506365776, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01500480342656374, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010387780144810677, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18597908318042755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0685216411948204, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06365900486707687, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06095382943749428, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05490431934595108, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03174234554171562, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.029438333585858345, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03740764036774635, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.034603167325258255, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03223595395684242, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02825031988322735, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027018483728170395, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01889747567474842, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.016541769728064537, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015306688845157623, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01500582043081522, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009487280622124672, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00804827269166708, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007908375933766365, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007364235818386078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007216283585876226, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005082744639366865, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005250627640634775, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004720207303762436, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0037041762843728065, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0685216411948204, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06793180108070374, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06288941204547882, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.059460680931806564, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05339454114437103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03112589381635189, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.028424248099327087, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03800031542778015, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03504443168640137, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03161217272281647, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027642572298645973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026586221531033516, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019116468727588654, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016738461330533028, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015015384182333946, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014578844420611858, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009608945809304714, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007879124023020267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007665649056434631, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007201718166470528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0069697340950369835, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005142101086676121, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005212864838540554, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0046380856074392796, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003603785764425993, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06793180108070374, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.25999733805656433, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24246487021446228, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23700159788131714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2140175849199295, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12166902422904968, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11514099687337875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1366496980190277, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12533994019031525, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1234307512640953, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1082298532128334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10308699309825897, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06923811882734299, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.059751398861408234, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.058089785277843475, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05770132690668106, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.034530121833086014, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.029351454228162766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02923741564154625, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.026564685627818108, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02646811306476593, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01763848029077053, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016945507377386093, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0169261135160923, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010085046291351318, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1366496980190277, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21367716789245605, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1357380598783493, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.099325992166996, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10173582285642624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08849339187145233, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04842331260442734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12159114331007004, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11183401197195053, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10045888274908066, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06016155704855919, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.062317974865436554, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06235292926430702, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05383988469839096, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04342059791088104, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.040624480694532394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03149149566888809, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023958537727594376, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02317967638373375, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019007159397006035, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017084695398807526, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.017683017998933792, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01801447384059429, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014632012695074081, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013377257622778416, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1357380598783493, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.22527219355106354, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.21149179339408875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.20731611549854279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18830697238445282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10585149377584457, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10084746032953262, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11710205674171448, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10841627418994904, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10718387365341187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09510506689548492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09078474342823029, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05974851921200752, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05202921852469444, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05088455602526665, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05060887336730957, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.029864590615034103, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02658376842737198, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.026509447023272514, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024470090866088867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.0244262907654047, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01592753641307354, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01663667894899845, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015559198334813118, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011809248477220535, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18830697238445282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2679446041584015, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2516327500343323, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2467477172613144, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22418084740638733, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12560147047042847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11967359483242035, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13878846168518066, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12858439981937408, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1271558403968811, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11285269260406494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10760447382926941, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07050463557243347, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06134799122810364, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06000747159123421, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05968998372554779, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03518108278512955, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03056471422314644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030475420877337456, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027941271662712097, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027896087616682053, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01823779195547104, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01794533059000969, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01779637672007084, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011304402723908424, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13878846168518066, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23404619097709656, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.213800311088562, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20646080374717712, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1872505396604538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10537239909172058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09865950047969818, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12006319314241409, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11079654842615128, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10757523030042648, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0939633920788765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08973429352045059, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06036654859781265, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05244596675038338, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05011706426739693, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.049559585750103, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030069950968027115, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02572138048708439, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025529230013489723, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023504791781306267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02318761497735977, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015755925327539444, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015669822692871094, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015007789246737957, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010242536664009094, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1872505396604538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08461999148130417, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07864335179328918, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07574872672557831, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06822697073221207, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03929680958390236, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03662056848406792, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04562907665967941, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04212800785899162, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0398898646235466, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03493109717965126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03331413120031357, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.023074625059962273, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02022360824048519, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.018955664709210396, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018641116097569466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011667918413877487, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010005838237702847, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009865591302514076, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009161571972072124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009015931747853756, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00634554959833622, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006532616913318634, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005959225352853537, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004683547653257847, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08461999148130417, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08028700947761536, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07442975044250488, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07112280279397964, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06386323273181915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03683856129646301, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.034043461084365845, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04353567212820053, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04031118005514145, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03740760684013367, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.032680854201316833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.031212320551276207, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02190183475613594, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019187910482287407, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01765427738428116, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.017286205664277077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010969736613333225, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00912235677242279, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008947193622589111, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00829428993165493, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008107493631541729, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00578333018347621, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005753234960138798, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005333669483661652, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0037726773880422115, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08028700947761536, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24553348124027252, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22907254099845886, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22369712591171265, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.201985165476799, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11476130038499832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10853435099124908, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12889806926250458, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11841470003128052, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11641135066747665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10214098542928696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09735865890979767, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06526672095060349, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05644265189766884, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05478404834866524, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05438164994120598, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03256765380501747, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027708951383829117, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027582554146647453, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.025094738230109215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024990273639559746, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016662167385220528, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016077805310487747, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0159747451543808, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009639556519687176, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12889806926250458, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.22049503028392792, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15785907208919525, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13351303339004517, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1168229877948761, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09738488495349884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07199256122112274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1220042034983635, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11129337549209595, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10299447923898697, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06342761963605881, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06481419503688812, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06242849677801132, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05398111417889595, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04775634780526161, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.046230338513851166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03202592954039574, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.026537423953413963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02602611482143402, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02072460763156414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01971004717051983, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01855645701289177, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01928115077316761, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.016813304275274277, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.015077263116836548, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15785907208919525, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.22812697291374207, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.21412451565265656, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.20987261831760406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19046688079833984, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10722913593053818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10211818665266037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11867547780275345, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10984958708286285, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10858698189258575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09627772122621536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09188210219144821, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.060563262552022934, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05271041393280029, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.051548637449741364, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.051265910267829895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.030266471207141876, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0269309151917696, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0268576480448246, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024773338809609413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024732695892453194, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01614457182586193, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016852760687470436, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015770016238093376, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011958105489611626, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19046688079833984, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2705872356891632, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2540739178657532, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24912717938423157, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22616082429885864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12688665091991425, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12087325751781464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14021416008472443, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12992404401302338, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12848421931266785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11393629014492035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10860641300678253, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07125131040811539, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06200116500258446, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06063401326537132, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.060313381254673004, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.035534486174583435, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030874695628881454, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030784854665398598, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02821449004113674, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028164684772491455, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01839950494468212, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018119316548109055, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017946727573871613, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011394830420613289, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14021416008472443, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23453757166862488, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21370743215084076, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20599310100078583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1870046854019165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10544802993535995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09846903383731842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12041696161031723, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11125265806913376, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10776399821043015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09396713227033615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08976385742425919, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06057688593864441, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.052715785801410675, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05018379166722298, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04958431050181389, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030260132625699043, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025813285261392593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025602396577596664, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023579278960824013, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023239480331540108, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015956448391079903, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01582586020231247, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01517115905880928, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010415931232273579, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1870046854019165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08315182477235794, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07734979689121246, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07435127347707748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06705624610185623, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03867914900183678, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03600520268082619, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.045172177255153656, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0416988804936409, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.039265863597393036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03440944850444794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03295688331127167, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.022884858772158623, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01991000398993492, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.018582869321107864, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018264446407556534, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011464787647128105, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009656400419771671, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009508395567536354, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008809692226350307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008655586279928684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006071149371564388, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006113989744335413, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005666048265993595, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004137376323342323, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08315182477235794, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08052656799554825, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07474487274885178, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0713665783405304, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06413136422634125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03710934892296791, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03427943214774132, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04392359405755997, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04070332646369934, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.037690166383981705, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.032961405813694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03148283809423447, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.022131001576781273, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019397037103772163, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.017802750691771507, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.017405355349183083, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011098094284534454, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009226984344422817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00902982335537672, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008388594724237919, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00819044653326273, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005860497709363699, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0058567943051457405, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00539200846105814, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0038740059826523066, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08052656799554825, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24247737228870392, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2261490523815155, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22072602808475494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19940203428268433, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11328107118606567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10714826732873917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12741975486278534, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11703775078058243, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11490515619516373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10086800903081894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09605681896209717, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0645354762673378, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05577880144119263, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05409005656838417, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.053700514137744904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.032188501209020615, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02734711393713951, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0272104162722826, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024754824116826057, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024663932621479034, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016452481970191002, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015843460336327553, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01575576886534691, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009468448348343372, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12741975486278534, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.20064380764961243, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14987100660800934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13132017850875854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10950076580047607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08750607818365097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06628087908029556, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10930477827787399, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10005304962396622, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09401577711105347, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06122971698641777, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.058175500482320786, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.055867671966552734, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04819600656628609, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.042537443339824677, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04111241549253464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.028179462999105453, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02305425889790058, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.022672642022371292, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018311647698283195, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017430268228054047, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015671560540795326, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0162260290235281, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013951146975159645, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012142093852162361, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14987100660800934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.22740976512432098, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2133917510509491, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.209146648645401, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18981504440307617, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10687626898288727, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1017516702413559, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11824734508991241, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10948661714792252, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10822170972824097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09590398520231247, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0914817526936531, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.060260798782110214, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05245053395628929, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.051276255398988724, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05099828541278839, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.030106298625469208, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.026615241542458534, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.026541387662291527, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024431394413113594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024388624355196953, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015946820378303528, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016398923471570015, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01556999608874321, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01134924404323101, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18981504440307617, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2699061632156372, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2533731162548065, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2484104335308075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22556759417057037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1266186386346817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12057652324438095, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13992147147655487, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1296551376581192, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1281924545764923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11364567279815674, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10835881531238556, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0710790604352951, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06186307221651077, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.060494840145111084, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.060172416269779205, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03546326234936714, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030811049044132233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030721358954906464, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02814776636660099, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02809913083910942, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01835600659251213, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018083203583955765, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017901863902807236, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011376517824828625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13992147147655487, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23039063811302185, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21002180874347687, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20244775712490082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18377679586410522, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10363800078630447, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09680822491645813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11844092607498169, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10934796929359436, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10589022934436798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09236139804124832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08824088424444199, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05965008586645126, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.051783669739961624, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04935069754719734, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.048763714730739594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029799170792102814, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025436867028474808, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02523503266274929, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023258859291672707, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02292594686150551, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01575392112135887, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015671683475375175, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014972097240388393, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.0104298684746027, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18377679586410522, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09037173539400101, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08392813801765442, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08100312203168869, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07296623289585114, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.041944075375795364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03915675729513168, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.048200372606515884, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04465389624238014, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.042571134865283966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03722761571407318, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.035480719059705734, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.024419430643320084, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.021331747993826866, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.020133517682552338, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.019842274487018585, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012219095602631569, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010473915375769138, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010349668562412262, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009541166014969349, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00941399298608303, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006456682458519936, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006611471530050039, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006083597894757986, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004530317150056362, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09037173539400101, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08906868100166321, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08258216083049774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07938363403081894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07126770913600922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04096636921167374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03807917609810829, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04767722636461258, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0441054068505764, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04158516228199005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03631153702735901, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03463909402489662, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.024048002436757088, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.021026797592639923, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01964533142745495, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01931125856935978, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.012049894779920578, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.010188890621066093, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.010033976286649704, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.009266314096748829, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.009110303595662117, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0063814520835876465, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006430154666304588, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005964791867882013, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00433326093479991, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08906868100166321, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23273029923439026, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21698641777038574, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21189039945602417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19104522466659546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1082913875579834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10244704782962799, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12142662703990936, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11170689016580582, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10984206944704056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0963892713189125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09170692414045334, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.061468884348869324, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.053211674094200134, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05166048929095268, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05131154879927635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030657751485705376, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026164527982473373, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02604256197810173, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023709066212177277, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02360786497592926, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015707753598690033, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015216559171676636, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015084909275174141, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009220561012625694, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19104522466659546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1963925063610077, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14992499351501465, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13184905052185059, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10646416991949081, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08651921898126602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06645894795656204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1091148853302002, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0993153527379036, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09212629497051239, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06013050302863121, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.057883888483047485, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05572093278169632, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0477922298014164, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04205844923853874, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04058825969696045, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.028031447902321815, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022615039721131325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.022149886935949326, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017767855897545815, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.016866758465766907, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015351908281445503, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015748821198940277, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013577701523900032, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011537613347172737, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14992499351501465, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21989712119102478, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2063048630952835, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.20217536389827728, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18347498774528503, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10328992456197739, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09834353625774384, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11423035711050034, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10582214593887329, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10459446161985397, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09272134304046631, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08846621215343475, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05825432017445564, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05074140802025795, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04960518330335617, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04934419319033623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.029100878164172173, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02584545873105526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025766856968402863, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02375582791864872, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023708928376436234, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015461660921573639, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016079183667898178, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015095897950232029, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01129249855875969, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18347498774528503, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2638148367404938, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24762767553329468, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24278774857521057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22043128311634064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12363746762275696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11776452511548996, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13652831315994263, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12660814821720123, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12521523237228394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11101659387350082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1058126837015152, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06935965269804001, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.060405053198337555, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0590677335858345, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.058749593794345856, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03458743914961815, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030067726969718933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029978778213262558, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027472084388136864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02742166630923748, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01786256581544876, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01763193868100643, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017418522387742996, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011064640246331692, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13652831315994263, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22682584822177887, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20657765865325928, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19901461899280548, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18038377165794373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.1019829660654068, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09515254944562912, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11691568791866302, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10776819288730621, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10422299802303314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09073430299758911, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08666213601827621, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05886344611644745, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.051064956933259964, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04858454689383507, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.047991540282964706, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029408684000372887, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025091061368584633, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02488137036561966, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.0229159127920866, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022576915100216866, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01559441164135933, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015537114813923836, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014798705466091633, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01041440386325121, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18038377165794373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06928656250238419, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06435751169919968, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06143553555011749, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05532058700919151, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03209598362445831, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.029653452336788177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03810775652527809, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0352623350918293, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03258645534515381, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.028532156720757484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027307406067848206, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019262151792645454, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01682441122829914, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015440084971487522, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015100411139428616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009655994363129139, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008071436546742916, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007907608523964882, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007369984872639179, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00719565199688077, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005121737252920866, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005227584857493639, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004691021516919136, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003588414518162608, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06928656250238419, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06557308882474899, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06078966706991196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05737433582544327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05149451270699501, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030012398958206177, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02735212817788124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03669917583465576, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03395913168787956, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.030496351420879364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02668391913175583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025598620995879173, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01843567192554474, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016189254820346832, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014406511560082436, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013970938511192799, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009224831126630306, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0074990238063037395, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007271414157003164, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006838454864919186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0065987673588097095, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004901523236185312, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004893462173640728, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004379174206405878, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032430498395115137, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06557308882474899, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22969703376293182, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21438704431056976, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20924437046051025, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1889447271823883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1071571633219719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1013546884059906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12106306850910187, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11090957373380661, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10867976397275925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09554426372051239, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09113427251577377, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.061367008835077286, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05285148695111275, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05117892473936081, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05074089765548706, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03060307539999485, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025893857702612877, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025748688727617264, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0234715286642313, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023341596126556396, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015651270747184753, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015046464279294014, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014925758354365826, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009012344293296337, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1889447271823883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17398780584335327, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1279108077287674, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10430942475795746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09259619563817978, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07379363477230072, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.051022786647081375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10340434312820435, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09494820237159729, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08192960917949677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0551016665995121, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05466547608375549, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05294876545667648, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0457138866186142, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03623480349779129, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03365182504057884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02659691497683525, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020033057779073715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019142616540193558, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01671242155134678, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015096370130777359, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014625879935920238, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015211203135550022, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011759338900446892, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011251324787735939, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17398780584335327, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21794810891151428, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2044394165277481, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.20032697916030884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1817290186882019, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10236945748329163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09745503216981888, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11324702203273773, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10489780455827713, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10365881770849228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09185490012168884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0876249298453331, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.057773590087890625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.050314050167798996, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04918179661035538, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.048919036984443665, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02888849377632141, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.025657745078206062, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02558593638241291, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.023587660863995552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023541759699583054, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015399173833429813, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01600557193160057, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015035819262266159, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011297985911369324, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1817290186882019, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.26353245973587036, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2473401129245758, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24247075617313385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.220112144947052, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12353751808404922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11764506250619888, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1365191489458084, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12649258971214294, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12509067356586456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11088717728853226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1056908667087555, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06936304271221161, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06035592034459114, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05901705101132393, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05870097130537033, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.034601036459207535, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0300572719424963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02996877208352089, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027458900585770607, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027408162131905556, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017927605658769608, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017646081745624542, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017483610659837723, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011101393029093742, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1365191489458084, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23108190298080444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21049237251281738, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20282627642154694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18389944732189178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10389873385429382, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09696632623672485, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11896800249814987, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10979053378105164, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10619840770959854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09253021329641342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08837078511714935, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05991670489311218, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05199884995818138, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0494609959423542, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04885069653391838, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029928971081972122, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02546466700732708, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025254515931010246, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023257369175553322, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022909948602318764, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01586388424038887, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015660172328352928, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015065700747072697, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010357429273426533, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18389944732189178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07942008227109909, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0738154873251915, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07096847891807556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06392820924520493, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.036850232630968094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03430338576436043, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04287872835993767, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03969799727201462, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0374152846634388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0327627919614315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03127877041697502, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021701456978917122, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01897081732749939, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017726119607686996, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.017421245574951172, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010871878825128078, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009264335967600346, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009123706258833408, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008457206189632416, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008313159458339214, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005774971563369036, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00593555299565196, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005393287632614374, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00411048112437129, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07942008227109909, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07822661101818085, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0725376307964325, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06932798027992249, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.062198977917432785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03592567890882492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03319098800420761, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04252796247601509, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03934689238667488, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03647791966795921, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03187895938754082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03045024536550045, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.021389726549386978, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01876099966466427, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.017240308225154877, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016871722415089607, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010731378570199013, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008958756923675537, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008774569258093834, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008153676055371761, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007965682074427605, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005695612169802189, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005723567679524422, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005252340342849493, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0038450518622994423, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07822661101818085, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23229850828647614, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21653762459754944, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21144087612628937, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1908463090658188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10832864046096802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10241611301898956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1218159943819046, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11200149357318878, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10988664627075195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09641904383897781, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09175647795200348, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.061789169907569885, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05334828421473503, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05171723663806915, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0513400174677372, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030775826424360275, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02619003690779209, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02605912648141384, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023709073662757874, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0236025620251894, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015759196132421494, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015239676460623741, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015107686631381512, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009197910316288471, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1908463090658188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21633702516555786, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16415832936763763, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1445450633764267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10964903980493546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09519212692975998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07423926144838333, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11866438388824463, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10881400853395462, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1016714945435524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06318793445825577, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06043698266148567, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06089558079838753, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05261920392513275, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04644486680626869, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.044894032180309296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.030853018164634705, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02533590979874134, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02487924136221409, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019625967368483543, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018673857674002647, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01727324165403843, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01798216812312603, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.015404213219881058, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013625400140881538, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16415832936763763, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.22015155851840973, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.206437885761261, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2022997885942459, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1834847331047058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10333675146102905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09836843609809875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11431635916233063, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10589251667261124, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10465826839208603, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09268687665462494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08836296200752258, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05822978913784027, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.050696007907390594, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04955529421567917, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.049281079322099686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02907981537282467, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.025668086484074593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02558923326432705, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.023538274690508842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023494383320212364, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015345352701842785, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01573723740875721, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014974215067923069, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010787691920995712, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1834847331047058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.26694491505622864, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2504608929157257, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24554970860481262, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22281160950660706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1251658946275711, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11917725950479507, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13828450441360474, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12817993760108948, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1267370730638504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11229507625102997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10702943056821823, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07026714086532593, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06116316840052605, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0598083958029747, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05948544666171074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03505222126841545, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030465319752693176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030374903231859207, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027820492163300514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027772510424256325, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01814895123243332, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017887338995933533, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017702441662549973, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011267435736954212, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13828450441360474, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23395507037639618, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21315079927444458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20530059933662415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1862965226173401, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10516615957021713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0981089323759079, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12049505859613419, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11125194281339645, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10752728581428528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09375706315040588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08950035274028778, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0605892576277256, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05265689641237259, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05003495514392853, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04940708726644516, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030221806839108467, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025695007294416428, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025476323440670967, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02347174659371376, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023110080510377884, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015897737815976143, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015719469636678696, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015074500814080238, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010262339375913143, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1862965226173401, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08291158080101013, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07698077708482742, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07405323535203934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06669729202985764, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03852871060371399, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.035835206508636475, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04465997964143753, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.041416313499212265, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0391276590526104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03418511524796486, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03257865086197853, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.022631453350186348, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.019807303324341774, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01853184774518013, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018223337829113007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011350519955158234, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00968211144208908, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009541671723127365, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008824989199638367, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008681388571858406, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006039151921868324, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006193190813064575, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005652009509503841, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0042882757261395454, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08291158080101013, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07674054801464081, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07113120704889297, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06788145005702972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06094131991267204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03529642894864082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.032590847462415695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04183473438024521, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0388108529150486, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.035879094153642654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.031321439892053604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0298599936068058, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02106497995555401, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.018452418968081474, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01695059984922409, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016590014100074768, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01056855358183384, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008781231008470058, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008607713505625725, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007984421215951443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0078006950207054615, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005592504981905222, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005583113990724087, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005148010794073343, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003707517869770527, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07674054801464081, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24485309422016144, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22833949327468872, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22306126356124878, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20132054388523102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11443718522787094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10815787315368652, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12866844236850739, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11813486367464066, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11602772027254105, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10180938988924026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09695398807525635, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06520494818687439, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.056319378316402435, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05463607609272003, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.054227788001298904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.032517921179533005, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027647031471133232, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027509594336152077, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.025031184777617455, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024923525750637054, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016648907214403152, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016047772020101547, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015952516347169876, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009639650583267212, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12866844236850739, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1918429434299469, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14134223759174347, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12144735455513, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10330682247877121, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0836239904165268, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.061524320393800735, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1055210754275322, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09740658849477768, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09020669758319855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05903696268796921, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0567023903131485, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0544184185564518, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.047376278787851334, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04116380959749222, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.039591234177351, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0280008465051651, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023068582639098167, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02260049432516098, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01884489692747593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01790255308151245, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.016425721347332, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01710079237818718, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014743583276867867, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013475456275045872, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1918429434299469, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21730990707874298, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2037656456232071, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1996404230594635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18114927411079407, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10208781808614731, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09716835618019104, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11297363042831421, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10461947321891785, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.103390634059906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09154593199491501, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08731669187545776, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05761725455522537, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05017051473259926, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04904712364077568, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.048776935786008835, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02882268838584423, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02555931732058525, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025487450882792473, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.023478655144572258, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023434815928339958, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015381287783384323, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01590707153081894, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015019585378468037, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011181178502738476, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18114927411079407, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.26485565304756165, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24852439761161804, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24362215399742126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22109490633010864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12418419122695923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11822183430194855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13722766935825348, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12717828154563904, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12576064467430115, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1113968938589096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10615482181310654, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0697188749909401, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06067798286676407, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05932687595486641, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05901060998439789, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03475981578230858, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030192948877811432, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030104735866189003, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027564695104956627, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027516966685652733, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017962295562028885, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017693836241960526, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01751943677663803, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011086465790867805, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13722766935825348, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2325676530599594, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21206894516944885, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20430858433246613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18535959720611572, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10461533814668655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09767013788223267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12016220390796661, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11075694113969803, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10696065425872803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09337358921766281, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08910802006721497, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06037513166666031, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05251474305987358, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.049879368394613266, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.049245744943618774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030095744878053665, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025841662660241127, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025613410398364067, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023679673671722412, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023316489532589912, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015819761902093887, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016122397035360336, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014972054399549961, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010933084413409233, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18535959720611572, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09262455254793167, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08607998490333557, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08333554118871689, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07506892085075378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.043056413531303406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04033954441547394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04916973039507866, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.045524176210165024, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04371391981840134, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.038222718983888626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.036373794078826904, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.024872006848454475, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.021712426096200943, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.020639855414628983, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.020381005480885506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012439442798495293, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010665484704077244, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010556790977716446, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009696630761027336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009588380344212055, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006522574461996555, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006604026537388563, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00618246104568243, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004413325805217028, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09262455254793167, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08844757825136185, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08199287205934525, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07888972759246826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07089968025684357, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.040694303810596466, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03789828345179558, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04729560390114784, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04376025125384331, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04134327173233032, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03608149290084839, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03437567874789238, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.023837590590119362, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02080550417304039, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0194914098829031, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.019180327653884888, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011934448964893818, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.010033683851361275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009890509769320488, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.009106779471039772, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008950444869697094, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006281236652284861, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006218465976417065, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00587426358833909, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004056238569319248, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08844757825136185, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23600377142429352, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22000010311603546, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2148592472076416, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1938590109348297, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10996869951486588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10401876270771027, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.123233862221241, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11347360908985138, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11155262589454651, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09787344187498093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09308432787656784, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.062335263937711716, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05405401438474655, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05245208740234375, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05206957459449768, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03108457289636135, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026538731530308723, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026404356583952904, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02402128465473652, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023917339742183685, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015905538573861122, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015375778079032898, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015298544429242611, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009214945137500763, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.123233862221241, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.20638629794120789, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16376084089279175, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14825370907783508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11728203296661377, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09184178709983826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07496685534715652, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11140146851539612, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10242914408445358, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09655171632766724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0663115605711937, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05864879488945007, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.056840766221284866, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04915560036897659, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04439768195152283, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04321817681193352, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02856367640197277, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023717302829027176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02334023453295231, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018963951617479324, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018322711810469627, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015611939132213593, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016114303842186928, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014136995188891888, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011751989834010601, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16376084089279175, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21806062757968903, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20447765290737152, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.20031610131263733, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18175934255123138, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10245029628276825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09750430285930634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11328869313001633, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1049647256731987, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10375896841287613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09189372509717941, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08763924241065979, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.057774171233177185, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0503494068980217, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.049224112182855606, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.048960261046886444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.028873058035969734, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02567451260983944, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025605209171772003, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.023597989231348038, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023552272468805313, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015344702638685703, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01601358875632286, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014985308982431889, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01129735168069601, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18175934255123138, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2645663917064667, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2482442557811737, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24335907399654388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22090591490268707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12405415624380112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11812370270490646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13707053661346436, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12703682482242584, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12561863660812378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11131902784109116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10607869923114777, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06963925063610077, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06061319634318352, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05927041918039322, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05895145609974861, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03473706170916557, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030182786285877228, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030095655471086502, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027568431571125984, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027516460046172142, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.0179761815816164, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017716355621814728, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01753256842494011, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011141099035739899, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13707053661346436, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23129622638225555, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21076087653636932, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2029232382774353, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1839940845966339, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10401640087366104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09703581035137177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11952615529298782, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11018451303243637, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10638337582349777, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09273900836706161, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08849404007196426, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.060138411819934845, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05223400890827179, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04958334565162659, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.048947613686323166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030071843415498734, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025664469227194786, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02543855644762516, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023490529507398605, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0231268759816885, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016025349497795105, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.0159857589751482, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015205939300358295, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010795610956847668, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1839940845966339, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09810536354780197, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09111624956130981, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0881781056523323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07939384877681732, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04555290937423706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.042669545859098434, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05225352570414543, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04815467819571495, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04625144600868225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.040430162101984024, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03858056291937828, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.026517437770962715, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.023066217079758644, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.021928364410996437, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.021651025861501694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013300152495503426, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.011473061516880989, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011362709105014801, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010465131141245365, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0103517547249794, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007080871146172285, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007304003462195396, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006720435805618763, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005134160164743662, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09810536354780197, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08781284093856812, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0813453197479248, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07812486588954926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0700584203004837, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04023921117186546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.037393417209386826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04683421924710274, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.043386854231357574, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.040887054055929184, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03564004227519035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.033925894647836685, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02356979064643383, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02061927691102028, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.019253075122833252, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.018910769373178482, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011800328269600868, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009883231483399868, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009734006598591805, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008953510783612728, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008799407631158829, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006172323599457741, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006110817193984985, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005751630291342735, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003940838389098644, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08781284093856812, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22098982334136963, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20578858256340027, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20086781680583954, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18102215230464935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10272770375013351, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0971008762717247, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11526770144701004, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10616036504507065, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10422906279563904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.091350257396698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08693407475948334, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0583379752933979, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05058196559548378, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04902276024222374, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04865521192550659, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0290914848446846, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024798132479190826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024677159264683723, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022443387657403946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.022340184077620506, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014879129827022552, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014409246854484081, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014276041649281979, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008646363392472267, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18102215230464935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21819429099559784, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17933273315429688, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.16577750444412231, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13655441999435425, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09812736511230469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0822930634021759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11742482334375381, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10714278370141983, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10213235020637512, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07459495216608047, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06590186059474945, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06014339253306389, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05159500613808632, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0474909171462059, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04649515822529793, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03031919151544571, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.025407791137695312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.025093713775277138, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021001890301704407, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.020501989871263504, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.016786951571702957, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01715736836194992, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.015502729453146458, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01268293708562851, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17933273315429688, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2163386046886444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20288537442684174, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19876748323440552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18033859133720398, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10162919014692307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09672512859106064, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11242272704839706, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10412129759788513, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10293081402778625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09116770327091217, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08692856132984161, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.057324595749378204, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.049933262169361115, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04881390184164047, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.048556290566921234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.028650449588894844, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02544974535703659, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025378897786140442, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.023382142186164856, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.0233375895768404, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015231835655868053, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01584731601178646, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014878622256219387, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01116210501641035, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18033859133720398, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2618418037891388, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2457316815853119, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2409295290708542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21863585710525513, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12274234741926193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11690160632133484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13560721278190613, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12569354474544525, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12430024147033691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11015846580266953, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10498008877038956, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06888844072818756, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05997262895107269, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05864379182457924, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05832966789603233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.034356553107500076, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029859596863389015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02977193519473076, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027267739176750183, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027221614494919777, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017762450501322746, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017514176666736603, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017324203625321388, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010999136604368687, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13560721278190613, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2285032570362091, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2079499512910843, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19996429979801178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18112260103225708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10274431854486465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09565860778093338, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11841995269060135, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10909611731767654, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10510867089033127, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09140938520431519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08720267564058304, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05959790572524071, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.051739271730184555, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.049003519117832184, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0483362078666687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029809754341840744, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025373170152306557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025137513875961304, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02319052442908287, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022808527573943138, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015872856602072716, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015847664326429367, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015018600970506668, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010721314698457718, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18112260103225708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09975361078977585, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09271114319562912, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08989700675010681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08097539097070694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.046280063688755035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.043440066277980804, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.052586860954761505, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04870448261499405, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.046985283493995667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04110720008611679, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03911265730857849, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.026641609147191048, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.023252243176102638, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.022179672494530678, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.021922223269939423, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013319841586053371, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01147009152919054, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011366172693669796, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010436142794787884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.010333869606256485, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0069840578362345695, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0071016401052474976, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006641826592385769, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004772564861923456, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09975361078977585, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09359356760978699, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08671722561120987, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.08366592973470688, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0751698762178421, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04294668510556221, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04011610895395279, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04947129264473915, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04584465175867081, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04365142807364464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03811383247375488, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03624965623021126, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.024919940158724785, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.021787121891975403, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0205575842410326, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.020256932824850082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.012463374994695187, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01053833868354559, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.010404452681541443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.009561141952872276, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.009422042407095432, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006505900993943214, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00644833967089653, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0061181047931313515, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004148039035499096, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09359356760978699, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23169736564159393, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21611008048057556, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21104104816913605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1905519664287567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10791301727294922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10209815949201584, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1207519993185997, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11138248443603516, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1094016507267952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0961177721619606, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09151662141084671, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06118876487016678, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.053041066974401474, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05149637162685394, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05112304538488388, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03052319586277008, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02605336159467697, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025931378826498985, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023615967482328415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02351902797818184, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015621068887412548, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0151404719799757, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015018452890217304, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009133785031735897, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1905519664287567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.20110569894313812, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17193548381328583, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1611424684524536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13123029470443726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0907684862613678, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0788780078291893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10821379721164703, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09929007291793823, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0937136858701706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07058905810117722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06377528607845306, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05537814274430275, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04783271998167038, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04393681138753891, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04299413412809372, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02783769741654396, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02361326850950718, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.023258088156580925, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019854756072163582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019409947097301483, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015320410951972008, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016083430498838425, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01415013987571001, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012013747356832027, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17193548381328583, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20805104076862335, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19512952864170074, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19122937321662903, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17346429824829102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09763497114181519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09294042736291885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10795102268457413, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10004784166812897, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.0988684892654419, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08762809634208679, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0835564136505127, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.054994843900203705, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.047969888895750046, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.046884261071681976, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04663342982530594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02748187631368637, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024445101618766785, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024372359737753868, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02247023768723011, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022427288815379143, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014562802389264107, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015227806754410267, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01421070285141468, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01072657573968172, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19122937321662903, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2516774535179138, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23622584342956543, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23154474794864655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21021099388599396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11788886785507202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11228613555431366, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13025552034378052, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12072734534740448, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11935614049434662, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10585890710353851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1008693128824234, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06615353375673294, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.057577621191740036, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05629584938287735, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05599117651581764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03298197314143181, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02864474058151245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028559384867548943, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026169268414378166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026118341833353043, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01703033410012722, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0167809147387743, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0166014451533556, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010508297011256218, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13025552034378052, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21837536990642548, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19837120175361633, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19033212959766388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17205935716629028, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09817980229854584, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09117616713047028, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11388717591762543, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10476808249950409, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10049687325954437, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08714072406291962, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.083123117685318, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0573626384139061, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.049796078354120255, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04697985574603081, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04628987982869148, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02872385084629059, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02462008036673069, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024366207420825958, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02253263257443905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02213972993195057, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015483335591852665, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015814200043678284, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014600163325667381, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011160551570355892, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19033212959766388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1012861505150795, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09422501921653748, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09139090776443481, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08247353881597519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04703204333782196, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04418942704796791, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05358074977993965, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04942498356103897, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0477355495095253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04188072681427002, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.039994481950998306, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02721245028078556, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02364588901400566, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.022599779069423676, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.022348154336214066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013632691465318203, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.011761006899178028, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011662764474749565, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010746760293841362, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.010649790056049824, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007205663248896599, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007381849456578493, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006874431855976582, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0050989254377782345, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1012861505150795, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08425618708133698, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07805578410625458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07502563297748566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06746470928192139, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03853360190987587, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0358702652156353, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04482902213931084, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.041577257215976715, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.039161715656518936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.034227948635816574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03263050690293312, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02259267307817936, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01977142132818699, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.018448926508426666, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.018122121691703796, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011303162202239037, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009503907524049282, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009358976036310196, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008643945679068565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008490635082125664, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005946632940322161, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005915111396461725, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005535765551030636, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0038696625269949436, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08425618708133698, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23139755427837372, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2158845067024231, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21115541458129883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19077661633491516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10760757327079773, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1019795835018158, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12047680467367172, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11075182259082794, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10909722745418549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09598182141780853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09150049835443497, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.060952357947826385, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05274881422519684, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.051306694746017456, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0509662888944149, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030413668602705002, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025954924523830414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02585010603070259, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023563168942928314, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02348133735358715, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015561679378151894, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01504332572221756, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0149812838062644, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009054425172507763, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19077661633491516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19261901080608368, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15726616978645325, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14375098049640656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12177909910678864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08464754372835159, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07074367254972458, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10454963892698288, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09643284976482391, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08980978280305862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06542235612869263, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06181829422712326, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05338612198829651, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.046288564801216125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04096751660108566, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03963011875748634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02681988663971424, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02195288985967636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.021527091041207314, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01836966536939144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017618536949157715, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014619193971157074, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015149136073887348, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012980332598090172, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011026844382286072, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19261901080608368, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20960991084575653, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19662438333034515, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19273591041564941, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1749703735113144, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09820391982793808, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09352749586105347, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10853392630815506, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10057420283555984, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09943289309740067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08818357437849045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08414860814809799, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05527571588754654, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04817263036966324, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04710697382688522, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04686245322227478, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02761649340391159, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024484364315867424, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024416819214820862, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022502390667796135, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022462770342826843, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014620899222791195, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015140432864427567, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01427872572094202, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010541386902332306, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19273591041564941, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.250805139541626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23550131916999817, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23091575503349304, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20968931913375854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11739455163478851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11183170974254608, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12964604794979095, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12015607208013535, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11887170374393463, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1054481565952301, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10054763406515121, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06582380086183548, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.057316072285175323, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05606665089726448, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05576815828680992, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.032826997339725494, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028546756133437157, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02846488729119301, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026094932109117508, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026049483567476273, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016966573894023895, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016742167994379997, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01655408926308155, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010520284995436668, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12964604794979095, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21985477209091187, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20089520514011383, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19352205097675323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1751886010169983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09907415509223938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09254945814609528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11374600231647491, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10491577535867691, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10120146721601486, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.088381826877594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08423765748739243, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05739055573940277, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04986683279275894, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04733790457248688, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04673667624592781, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028739135712385178, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024707110598683357, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024491284042596817, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022671421989798546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022329574450850487, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015429943799972534, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015674743801355362, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014653335325419903, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010950966738164425, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19352205097675323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09462461620569229, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08796553313732147, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08512423187494278, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0767783597111702, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04385574534535408, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04108124226331711, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05045924335718155, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.046407341957092285, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04453759267926216, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03902394697070122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03730711340904236, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.025566859170794487, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.022241337224841118, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.021107688546180725, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.020833682268857956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012845871970057487, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.011071671731770039, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010955555364489555, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010124445892870426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01000690646469593, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006882001645863056, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007084808778017759, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006526730488985777, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005006716586649418, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09462461620569229, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08280526101589203, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07652969658374786, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0733509212732315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06577526032924652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03773841634392738, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03494786471128464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04423541575670242, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04095403477549553, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.038349200040102005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03343239426612854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.031801462173461914, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.022238323464989662, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019493592903017998, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01807248964905739, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01772657223045826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011130303144454956, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009339429438114166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009181290864944458, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008483552373945713, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008305741474032402, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005860780831426382, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005891557317227125, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005422221962362528, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003910096827894449, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08280526101589203, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.228732168674469, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21344824135303497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20856761932373047, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1883256435394287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10648215562105179, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10079364478588104, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1191219761967659, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10975662618875504, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10795355588197708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09492339938879013, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09037710726261139, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.060333557426929474, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05225689709186554, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0507904551923275, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.050429150462150574, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030058279633522034, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025679336860775948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025564143434166908, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023277854546904564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02318635582923889, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015392061322927475, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01486077532172203, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014819175004959106, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008899238891899586, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1883256435394287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16617529094219208, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1400202214717865, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12981700897216797, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1042359471321106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07479451596736908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06401912868022919, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09127447009086609, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08290421962738037, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07737275213003159, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.056842539459466934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.052136823534965515, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04670422896742821, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.040197789669036865, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03660101443529129, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.035702645778656006, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024005474522709846, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020167358219623566, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0198274664580822, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016956770792603493, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01651592180132866, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01400733646005392, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014333085156977177, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012986576184630394, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011217969469726086, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16617529094219208, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.196462482213974, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18445920944213867, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18084198236465454, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16419589519500732, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09199660271406174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0876590833067894, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10169628262519836, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09418428689241409, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09312645345926285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08267182111740112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07895750552415848, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.051843322813510895, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.045115407556295395, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04412873089313507, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04389427974820137, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.025919929146766663, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022949647158384323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.022884653881192207, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.021107535809278488, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.021069476380944252, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013790756464004517, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014211050234735012, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0134727843105793, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009920847602188587, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18445920944213867, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24278615415096283, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22811631858348846, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22369663417339325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20332396030426025, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11362479627132416, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10829729586839676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12550432980060577, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11629223078489304, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11500390619039536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10217013955116272, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09749779105186462, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06378407776355743, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.055476315319538116, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0542755201458931, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05399683862924576, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03183046728372574, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.027703121304512024, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027627920731902122, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025361623615026474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025320036336779594, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01656617969274521, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01635347306728363, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016175931319594383, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01043590996414423, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12550432980060577, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21506232023239136, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1971200704574585, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19013246893882751, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17242386937141418, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09691076725721359, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0907265767455101, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11134936660528183, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.102646604180336, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09898938983678818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08678554743528366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08285144716501236, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05616118386387825, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.048718709498643875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.046235159039497375, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0456390343606472, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028058992698788643, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.023957274854183197, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023737408220767975, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021999549120664597, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021663794293999672, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01489320956170559, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014958449639379978, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014097354374825954, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.0101547259837389, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19013246893882751, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09908567368984222, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0923178568482399, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08957941085100174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0809604674577713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0459568165242672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04321935772895813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.052436672151088715, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04828687012195587, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04661915823817253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04097968712449074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.039234839379787445, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.026564227417111397, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02308608964085579, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0220497939735651, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02180199697613716, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013303149491548538, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.011434856802225113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011334113776683807, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010447822511196136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.010350285097956657, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007013052701950073, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007118752226233482, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006685327272862196, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004847628530114889, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09908567368984222, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08698622137308121, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08064129948616028, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07760191708803177, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06979009509086609, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03972746431827545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03698628395795822, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04626781865954399, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04269515722990036, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04036428779363632, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.035314273089170456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03373023867607117, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02331375703215599, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.020348062738776207, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.019020168110728264, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.018702764064073563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011658918112516403, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009825684130191803, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00967361032962799, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008944573812186718, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008789177983999252, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0061347526498138905, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006142502650618553, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005721075925976038, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004072700627148151, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08698622137308121, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2345874011516571, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21929147839546204, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2144932597875595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19417402148246765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1091931164264679, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10362895578145981, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12189948558807373, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11248797178268433, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11070059984922409, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09760180115699768, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09308332949876785, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06173134595155716, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05356919765472412, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05213030427694321, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05178210139274597, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030794167891144753, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026397183537483215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026287658140063286, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023986216634511948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023915641009807587, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015780484303832054, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015320749022066593, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015241747722029686, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009288710542023182, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12189948558807373, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17306700348854065, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14553944766521454, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13647820055484772, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10758620500564575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0769137293100357, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06679671257734299, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0927661582827568, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08381712436676025, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08058144897222519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05969010666012764, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05134854093194008, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04716341570019722, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.040547795593738556, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03743795305490494, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03667620196938515, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024478338658809662, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020357782021164894, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020153585821390152, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017200693488121033, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.016872374340891838, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014451645314693451, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01408771239221096, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01353986281901598, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01079826895147562, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17306700348854065, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1936863213777542, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18187573552131653, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1783452183008194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16207438707351685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09061628580093384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08639194816350937, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10018840432167053, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09277136623859406, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09174015372991562, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08149832487106323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07794699817895889, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.051058728247880936, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.044416677206754684, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.043453194200992584, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04322289675474167, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.025521570816636086, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022566353902220726, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02250208519399166, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020753972232341766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02072218433022499, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013554131612181664, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013920111581683159, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013244858011603355, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009661057963967323, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1936863213777542, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24214990437030792, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22760041058063507, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22324307262897491, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20302614569664001, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1132500022649765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10801521688699722, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12509198486804962, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11586851626634598, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11464686691761017, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10193825513124466, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09735292941331863, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06356505304574966, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.055283304303884506, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05409718677401543, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.053814928978681564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03170325607061386, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.027590468525886536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027514366433024406, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025269264355301857, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02522747591137886, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01645483262836933, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016253545880317688, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016066892072558403, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010327289812266827, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12509198486804962, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20776623487472534, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19029128551483154, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18329748511314392, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16654449701309204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09361598640680313, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08746004104614258, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1078827977180481, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09956364333629608, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09570086747407913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08386476337909698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08017390966415405, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05437391996383667, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04719848185777664, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.044621631503105164, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04400407522916794, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027137424796819687, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02302110567688942, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022789740934967995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021123772487044334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0207652747631073, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014301062561571598, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014270656742155552, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013470938429236412, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009478472173213959, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19029128551483154, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0861949473619461, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08018244057893753, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07744447886943817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06997035443782806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03983473777770996, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03729020059108734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04593738168478012, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0424274280667305, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04044440761208534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03553696349263191, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0339946448802948, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02325247786939144, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.020282890647649765, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019150374457240105, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018876591697335243, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011680380441248417, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010020282119512558, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009903074242174625, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009177405387163162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009054966270923615, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006252671591937542, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006401943042874336, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0059039960615336895, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004484198521822691, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0861949473619461, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07745537161827087, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07175671309232712, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06856904923915863, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06173338741064072, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03527631238102913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.032640088349580765, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.041720982640981674, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03857363760471344, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.035863280296325684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03139229491353035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.030033696442842484, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02096611075103283, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01837080530822277, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016885505989193916, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016536913812160492, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010491923429071903, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008733409456908703, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008565856143832207, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00795982126146555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007784049492329359, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005553394556045532, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005528636742383242, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0051097325049340725, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003658560337498784, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07745537161827087, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21002346277236938, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19600307941436768, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19147400557994843, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17299994826316833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09720510244369507, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09202754497528076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10891076177358627, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10021732747554779, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09852532297372818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08678543567657471, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08271335810422897, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.055062174797058105, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04769434034824371, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04631751403212547, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.045986779034137726, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027432303875684738, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023428266867995262, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023315606638789177, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021276291459798813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021196063607931137, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014044092036783695, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01358866784721613, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01351969875395298, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008172399364411831, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19147400557994843, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17696800827980042, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1452445536851883, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.130497545003891, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1007397323846817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07908392697572708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0650944709777832, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10055577009916306, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09249962866306305, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08277825266122818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05796503648161888, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05368902161717415, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05138949677348137, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04444223642349243, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.038381483405828476, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03683217614889145, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025805260986089706, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020674820989370346, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02003946155309677, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016784977167844772, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015918027609586716, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014085698872804642, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014550537802278996, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012223522178828716, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010591056197881699, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17696800827980042, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19067610800266266, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1791958510875702, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17574277520179749, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1597749888896942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08914868533611298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08501330018043518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09849587082862854, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09124334901571274, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09024044126272202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0802542194724083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0766516625881195, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05009554326534271, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04363928362727165, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04269903153181076, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04247436672449112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.025023551657795906, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02210417576134205, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.022042540833353996, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020329097285866737, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02029329538345337, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013185563497245312, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013530664145946503, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012880702503025532, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009259868413209915, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19067610800266266, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23884910345077515, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22462904453277588, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22038856148719788, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2005327194929123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11168244481086731, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10656407475471497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.123334139585495, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1142306700348854, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11302758008241653, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10060146450996399, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09610653668642044, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06262511014938354, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05447520688176155, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.053322236984968185, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.053050775080919266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.031247111037373543, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.027171416208148003, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027098607271909714, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024901146069169044, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024863088503479958, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.0162216667085886, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01597008854150772, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01584242656826973, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010098457336425781, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.123334139585495, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2089615762233734, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19136321544647217, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18436063826084137, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16745203733444214, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09411489963531494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08795695006847382, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10869193822145462, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10006637126207352, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.0962735265493393, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08433641493320465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08062976598739624, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05474746972322464, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0475139245390892, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04492713138461113, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0443049818277359, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02740694023668766, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.023306384682655334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023077767342329025, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02141745761036873, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021062353625893593, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014629168435931206, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014619986526668072, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013807715848088264, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009945519268512726, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19136321544647217, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.10552220791578293, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09839101135730743, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09558063745498657, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08645229786634445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.048890385776758194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04602792486548424, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.055622782558202744, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05129609629511833, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04958636686205864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04368159547448158, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.041819650679826736, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.028204424306750298, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.024520423263311386, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.023456968367099762, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.023205868899822235, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0141224330291152, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.012202008627355099, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.012099001556634903, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.011173996143043041, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01107446663081646, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007481955457478762, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007637683767825365, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00714521948248148, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005268254317343235, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.10552220791578293, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08708278089761734, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0809055045247078, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07773277908563614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06998641788959503, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03973597288131714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.036979977041482925, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04643359035253525, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.042876824736595154, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0403263084590435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.035413213074207306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03382924944162369, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.023356836289167404, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.020401693880558014, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01899527944624424, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.018647488206624985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011680003255605698, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009761697612702847, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009599468670785427, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008892394602298737, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008724688552320004, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006124573759734631, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006050326861441135, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005688515957444906, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003924437798559666, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08708278089761734, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22527135908603668, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2104957550764084, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2059773951768875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18646785616874695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10431074351072311, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09900327026844025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11657765507698059, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10730227828025818, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10565893352031708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09333977848291397, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08927671611309052, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.058991312980651855, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05104045197367668, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.049692727625370026, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04937348887324333, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.029401589184999466, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02514275163412094, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025037189945578575, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02288801781833172, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.022811438888311386, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015061154961585999, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014572611078619957, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014514745213091373, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008782872930169106, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18646785616874695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17680642008781433, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1527620106935501, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14515168964862823, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12372659891843796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07901392132043839, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07009968161582947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09275619685649872, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08441756665706635, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0817856416106224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06410055607557297, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05909866839647293, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04697343334555626, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04063856229186058, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03822435438632965, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.037636298686265945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024065406993031502, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020498178899288177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020347431302070618, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01770995371043682, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017468493431806564, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013728165999054909, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013741455040872097, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01301777083426714, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010314340703189373, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17680642008781433, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18002483248710632, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.16918182373046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.16592486202716827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1507447063922882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08406507223844528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08015891164541245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09302996844053268, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08606600761413574, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08510159701108932, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07566548883914948, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07228823751211166, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.047296635806560516, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04114305227994919, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04024151712656021, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.040026504546403885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.023645544424653053, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02079244703054428, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.020735586062073708, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019109755754470825, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.019076984375715256, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012513796798884869, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012668710201978683, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012221684679389, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008601098321378231, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18002483248710632, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23672166466712952, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22268763184547424, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2184663712978363, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1987963765859604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11062337458133698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10555572062730789, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12216685712337494, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11318513751029968, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11196094006299973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09968121349811554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09519604593515396, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.062046121805906296, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05398549139499664, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05282919853925705, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05255667865276337, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.030942745506763458, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.026922430843114853, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02684452198445797, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02467244304716587, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024634618312120438, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016042960807681084, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015834005549550056, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015664033591747284, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010019343346357346, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12216685712337494, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20223653316497803, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18648135662078857, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18029387295246124, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16410037875175476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09137944132089615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08586907386779785, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10458335280418396, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09664187580347061, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09325084835290909, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0822954848408699, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07873009145259857, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.052694469690322876, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04580515995621681, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.043512001633644104, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04296461492776871, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.026270398870110512, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02234163135290146, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022136280313134193, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02055950090289116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020243139937520027, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013744582422077656, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013661932200193405, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013002256862819195, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.008904038928449154, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18648135662078857, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08967824280261993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08361253142356873, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08090949058532715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07318480312824249, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04156603664159775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03899736702442169, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04763798415660858, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.044050123542547226, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.042164549231529236, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03711896762251854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.035504069179296494, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02414034865796566, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.021061204373836517, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019967684522271156, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01970399171113968, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01209320593625307, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01042245514690876, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010307183489203453, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009549579583108425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009436708875000477, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006424717605113983, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006611640565097332, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0060883089900016785, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004596329759806395, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08967824280261993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08168306201696396, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07596132904291153, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07279129326343536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06564170867204666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03738783299922943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03472820669412613, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04415464773774147, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.040643010288476944, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03795154392719269, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03334098309278488, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03193320333957672, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.022228194400668144, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019358742982149124, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.017921673133969307, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.017570942640304565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011138195171952248, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009269895032048225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009103294461965561, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008467436768114567, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008293996565043926, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005885663442313671, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0058584315702319145, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005440068431198597, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0038842337671667337, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08168306201696396, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23531512916088104, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2203684002161026, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21559594571590424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19524531066417694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10975965112447739, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10413990169763565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1225351095199585, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11292119324207306, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.111115463078022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09818539768457413, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09373611956834793, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06204160302877426, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05378326028585434, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05234115570783615, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05199194326996803, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03095826506614685, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026485467329621315, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026372741907835007, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024093899875879288, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024017486721277237, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015858912840485573, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01534956507384777, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01530674658715725, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00925024226307869, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1225351095199585, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1723974049091339, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13447824120521545, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1216607391834259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10275672376155853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07600607722997665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06159301474690437, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09256132692098618, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08421221375465393, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0803755521774292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05516888573765755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.052307091653347015, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0472133569419384, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04067835584282875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03701642155647278, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03611133620142937, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02408999763429165, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020179787650704384, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01994079165160656, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016411298885941505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015932301059365273, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013833796605467796, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014113480225205421, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012781083583831787, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010813720524311066, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1723974049091339, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18308782577514648, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17209121584892273, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1687740832567215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.15354901552200317, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08553817123174667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08156738430261612, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09466264396905899, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08756303042173386, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08657313138246536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07702741026878357, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07366511970758438, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04815089702606201, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04183957353234291, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.040930382907390594, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.040716059505939484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02406676858663559, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.021119069308042526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02105739153921604, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01940818876028061, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.019376715645194054, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012740127742290497, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012820805422961712, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01245080679655075, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008646038360893726, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18308782577514648, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24032416939735413, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22613120079040527, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2218659669160843, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2019529789686203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11238440871238708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10724852979183197, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12419681996107101, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1149776354432106, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11373578011989594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10130783170461655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09687379002571106, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06312894821166992, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.054862577468156815, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05368399992585182, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05340948700904846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03150235116481781, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02740868739783764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027331745252013206, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025146588683128357, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025102825835347176, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01642114669084549, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016193939372897148, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01603211835026741, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010355761274695396, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12419681996107101, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20713259279727936, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19149592518806458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18525844812393188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1686686873435974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09361529350280762, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08813785016536713, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10716764628887177, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09895287454128265, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09549771249294281, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08452508598566055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08090189099311829, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05396990478038788, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04689370095729828, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04458273947238922, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04403107985854149, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.026914920657873154, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022906139492988586, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022696388885378838, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021124014630913734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020806316286325455, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014113240875303745, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014019171707332134, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013369721360504627, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.00916240643709898, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19149592518806458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08339555561542511, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07754859328269958, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07470212876796722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06748999655246735, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03855251148343086, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.035977113991975784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04462563991546631, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04134591668844223, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03914701193571091, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03436160087585449, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03286435082554817, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.022567961364984512, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01973947137594223, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01849738508462906, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018197493627667427, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011303291656076908, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009612706489861012, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009477809071540833, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008782961405813694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008642012253403664, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0059726666659116745, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006079018581658602, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0055937389843165874, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004134935326874256, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08339555561542511, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07553400099277496, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07018423080444336, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06667976081371307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06018475815653801, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.034448038786649704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.031717121601104736, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.041611745953559875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03836653754115105, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03498350828886032, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.030786562711000443, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.029636891558766365, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020924074575304985, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0182233527302742, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016522208228707314, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016089068725705147, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010482213459908962, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008537067100405693, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0083268778398633, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007800000719726086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007576650474220514, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005533063318580389, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005450663156807423, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005021111574023962, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0035625144373625517, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07553400099277496, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22858749330043793, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2135012298822403, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20874813199043274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18877287209033966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10588336735963821, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10028737783432007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11856045573949814, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10898354649543762, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10737928748130798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09455014765262604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09022144973278046, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05995350331068039, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05185488983988762, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.050424184650182724, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05010296404361725, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.029885029420256615, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02545476146042347, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025356190279126167, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023111617192626, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02302665077149868, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015262224711477757, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01467728242278099, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014674346894025803, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008719345554709435, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18877287209033966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17119236290454865, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14669716358184814, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1341797113418579, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11260084807872772, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0755668506026268, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06420537829399109, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09881921112537384, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0891944169998169, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07934987545013428, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06237088143825531, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05678917095065117, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05018486827611923, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04329995810985565, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03725283965468407, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0356912687420845, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02575693279504776, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02099313586950302, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020390378311276436, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018710361793637276, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017893996089696884, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0151098957285285, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015755634754896164, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013409341685473919, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012504970654845238, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17119236290454865, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18619687855243683, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1749664843082428, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17163856327533722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1561395674943924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08696785569190979, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08294250071048737, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09621824324131012, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.0890364795923233, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08805198967456818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07832317054271698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07491976022720337, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04890495538711548, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04253854975104332, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.041613221168518066, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.041395992040634155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.024429775774478912, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.021449575200676918, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.021388379856944084, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019708111882209778, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.019673924893140793, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012865567579865456, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012987632304430008, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012569154612720013, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008720140904188156, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18619687855243683, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24319712817668915, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22889286279678345, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22458267211914062, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20447991788387299, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11378596723079681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10857978463172913, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12567934393882751, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11643067002296448, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11514860391616821, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10257662087678909, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0980619341135025, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06385334581136703, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05553751438856125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05434834212064743, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05406690388917923, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03185223788022995, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02771913632750511, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02764158695936203, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025420090183615685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025378482416272163, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016551876440644264, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016336150467395782, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016164714470505714, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01038377359509468, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12567934393882751, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21175268292427063, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.196400448679924, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1903505176305771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17346012592315674, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0958048552274704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0904960036277771, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10925259441137314, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10093610733747482, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09759391099214554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08676136285066605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08303233236074448, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05493338033556938, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.047771383076906204, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04555176943540573, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04502453655004501, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027379723265767097, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.023266803473234177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023065173998475075, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02147528901696205, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02117268554866314, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014253849163651466, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014013939537107944, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013537110760807991, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.008894510567188263, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1903505176305771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08912377804517746, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08329091966152191, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08039683848619461, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07288969308137894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.041362371295690536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03874028101563454, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04799869656562805, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04429510235786438, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.041929543018341064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03709906339645386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03561558201909065, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.024308837950229645, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.021210424602031708, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019917314872145653, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.019599447026848793, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012269536033272743, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010472811758518219, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010330191813409328, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009643022902309895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009496374055743217, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00666093360632658, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006775842979550362, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006281038746237755, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004808458499610424, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08912377804517746, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0835810974240303, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07790596783161163, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07469887286424637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0673726350069046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03835534676909447, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.035607531666755676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04526767507195473, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04186573252081871, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03890334069728851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03431151062250137, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03286588191986084, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.022748764604330063, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019941121339797974, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01835881918668747, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.017975950613617897, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011386947706341743, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009484522044658661, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00929694902151823, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008678697980940342, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008486679755151272, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006027826573699713, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005987942684441805, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005568781867623329, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003938059322535992, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0835810974240303, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23971039056777954, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2250884771347046, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22044581174850464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20035666227340698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11198192834854126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1065157875418663, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12497279793024063, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11504185199737549, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1133570522069931, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1006544753909111, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09629359096288681, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06340274214744568, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05482346564531326, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05341954901814461, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05308138579130173, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.031614042818546295, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027035560458898544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02692977339029312, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024680549278855324, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024616124108433723, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016207614913582802, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015673821792006493, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01563054881989956, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009475798346102238, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12497279793024063, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17424607276916504, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14138783514499664, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13044893741607666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10563773661851883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07650002092123032, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06399981677532196, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09372115135192871, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0842050313949585, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08099601417779922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05706047639250755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.052007194608449936, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0473751574754715, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.041091568768024445, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03768572956323624, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03684421628713608, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024645093828439713, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02112303115427494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020942447707057, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01774664595723152, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017351342365145683, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014686595648527145, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015287278220057487, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013740112073719501, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012286169454455376, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17424607276916504, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18158002197742462, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17079561948776245, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1675693690776825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.152528315782547, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08483926951885223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08094494044780731, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09383603930473328, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08684853464365005, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08587672561407089, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07645740360021591, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07312680780887604, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04767429828643799, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04148448631167412, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04059029370546341, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.040376853197813034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02381621114909649, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.020904509350657463, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.020844032987952232, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019218111410737038, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01918276958167553, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012524147517979145, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012635299935936928, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012233303859829903, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008453525602817535, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18158002197742462, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23851318657398224, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22456935048103333, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2203652709722519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20074832439422607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11156927049160004, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1065215989947319, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12327075004577637, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11414630711078644, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1129070371389389, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10067307949066162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09626610577106476, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06261985749006271, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05445193499326706, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05329160764813423, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05301651731133461, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0312502458691597, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.027173196896910667, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027096601203083992, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024933774024248123, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02489345148205757, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016255099326372147, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016002845019102097, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015877151861786842, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010165884159505367, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12327075004577637, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20428438484668732, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1895858496427536, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18350276350975037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1671159863471985, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09253406524658203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08731044828891754, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10614998638629913, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09785550832748413, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09425638616085052, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08378764986991882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08018998056650162, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05337978154420853, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.046409182250499725, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04409194365143776, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04353700950741768, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.026722241193056107, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022695256397128105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02247997187077999, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02098035253584385, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02066960372030735, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014138918370008469, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013943950645625591, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013401038944721222, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009197269566357136, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1895858496427536, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07730567455291748, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.072027787566185, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06913799047470093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06256327033042908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03582712262868881, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.033314235508441925, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04225245863199234, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03886004537343979, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03636634722352028, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03203577920794487, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03080795332789421, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021376490592956543, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01862383633852005, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017256228253245354, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01692173257470131, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010774138383567333, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009067104198038578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008910898119211197, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00832456350326538, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008162888698279858, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005818508565425873, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005894920323044062, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005416171625256538, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004145848099142313, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07730567455291748, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06820952892303467, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06336670368909836, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.059969738125801086, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0540291927754879, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03108450211584568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02847316302359104, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.037975724786520004, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.035005420446395874, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.031573351472616196, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027791187167167664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02675252966582775, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019076386466622353, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016665203496813774, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014905950985848904, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014480525627732277, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009545592591166496, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007729346863925457, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007512702606618404, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007069712039083242, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0068361032754182816, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005062241107225418, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004995432216674089, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0045491233468055725, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032918169163167477, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06820952892303467, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22840777039527893, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21342726051807404, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20862635970115662, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18886636197566986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1060848981142044, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10058378428220749, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11899147927761078, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1093045026063919, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10755927860736847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09475443512201309, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09046173840761185, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06020183116197586, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0520310252904892, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05060463771224022, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.050257742404937744, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030031904578208923, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025609787553548813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025501040741801262, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02326519601047039, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023178238421678543, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015380671247839928, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014865679666399956, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014777596108615398, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008974799886345863, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18886636197566986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16997699439525604, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13547182083129883, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12193657457828522, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10263754427433014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07209080457687378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05934489518404007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09578274190425873, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08530634641647339, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0788361132144928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.056729648262262344, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05383530259132385, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.048399731516838074, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04164767637848854, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03581999987363815, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.034292470663785934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025178974494338036, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02060740813612938, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020217927172780037, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01803409866988659, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01723591797053814, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015272865071892738, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015823883935809135, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013659331947565079, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012895301915705204, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16997699439525604, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1846749484539032, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1737717092037201, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17044636607170105, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1552024781703949, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08633434772491455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08237208425998688, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.0954434797167778, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08833451569080353, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08736057579517365, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07782105356454849, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07446534931659698, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0484846793115139, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04221215099096298, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04130455106496811, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04109155014157295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02422947995364666, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.021296877413988113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.021235831081867218, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019586218520998955, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01955253817141056, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012752185575664043, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012904340401291847, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012461184523999691, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008676238358020782, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1846749484539032, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2400277853012085, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22602875530719757, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22181488573551178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20212554931640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11231201142072678, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10721077024936676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12404531240463257, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11488677561283112, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11363635957241058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10132472962141037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09693045914173126, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06301753968000412, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.054805804044008255, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05364788696169853, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.053370289504528046, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.031435105949640274, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.027362961322069168, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027287913486361504, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025115152820944786, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02507646009325981, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016322195529937744, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016128910705447197, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015941817313432693, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010263078846037388, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12404531240463257, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2091159075498581, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19442646205425262, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18856912851333618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17183807492256165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0947457104921341, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08964825421571732, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10780491679906845, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.099683478474617, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09640811383724213, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0859135314822197, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08217300474643707, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.054210446774959564, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04721980169415474, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.045070942491292953, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04455484822392464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027048859745264053, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02305252104997635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022854121401906013, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.021309614181518555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021016523241996765, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014105780981481075, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013927586376667023, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013413854874670506, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.008921755477786064, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18856912851333618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06601187586784363, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06150782108306885, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05869380757212639, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05308925360441208, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03051699884235859, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.028230486437678337, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0363045260310173, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03361644223332405, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03097708150744438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.027292734012007713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.026222623884677887, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018331827595829964, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.016061825677752495, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01469537615776062, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0143560990691185, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009215681813657284, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0077218725346028805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007561000995337963, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007092798128724098, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006923471577465534, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004961599130183458, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005061684176325798, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004569980315864086, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0035418523475527763, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06601187586784363, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.058822471648454666, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.054747965186834335, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.051571860909461975, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.046570658683776855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02695174515247345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.024565499275922775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03335275873541832, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.030671175569295883, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02734503522515297, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.024126781150698662, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02332414872944355, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01676575466990471, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.014632384292781353, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.012952215038239956, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.012524950318038464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00841123890131712, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006743319798260927, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006528504192829132, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006189887411892414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005963465664535761, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004468908067792654, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00442897342145443, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003984448965638876, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002955311443656683, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.058822471648454666, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21986961364746094, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20571957528591156, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20109321177005768, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.182292640209198, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10251206159591675, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09712990373373032, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11502984166145325, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1057685911655426, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1039566770195961, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09169197082519531, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08749668300151825, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.058239180594682693, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0503787100315094, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04890873283147812, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04854985326528549, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.029052171856164932, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02471413090825081, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024592839181423187, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02245347574353218, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02235693298280239, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014841340482234955, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014296658337116241, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014234928414225578, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008504004217684269, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.182292640209198, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14753855764865875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11458196491003036, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10218130052089691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08373154699802399, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06489565968513489, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.051374491304159164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08150951564311981, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07420574873685837, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0691024586558342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04614109545946121, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04542043060064316, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04192621633410454, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03618841990828514, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03209209069609642, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.031059281900525093, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.021712232381105423, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.018255608156323433, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017950834706425667, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015114622190594673, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014544254168868065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013073796406388283, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013672334142029285, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01196750532835722, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011116414330899715, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14753855764865875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18929962813854218, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17812837660312653, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17475955188274384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1592053472995758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0885193943977356, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0844850093126297, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09805265069007874, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09057614207267761, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08957131952047348, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07984283566474915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07657304406166077, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04990823566913605, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04331396147608757, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04239300265908241, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.042169876396656036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.024973440915346146, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.021917805075645447, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.021859755739569664, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02018097974359989, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.020149653777480125, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013290312141180038, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013376139104366302, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01300151739269495, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009119998663663864, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18929962813854218, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24162837862968445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22756452858448029, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22331435978412628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2035558968782425, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11309263110160828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10797680169343948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12503549456596375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11565835773944855, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1144014298915863, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10208317637443542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09777715057134628, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06360555440187454, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05521545559167862, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05404765531420708, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.053767815232276917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03174690529704094, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.027617350220680237, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027539703994989395, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02536889538168907, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025330234318971634, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016618771478533745, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0163637176156044, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01624384894967079, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010529122315347195, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12503549456596375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21058440208435059, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1962335705757141, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19028612971305847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1735183745622635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0954986959695816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09042421728372574, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10928420722484589, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10073710978031158, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09713156521320343, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08683162927627563, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08313388377428055, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05492236837744713, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04774419963359833, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04546770080924034, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.044925957918167114, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027461757883429527, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02333204261958599, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023116491734981537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02162889577448368, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021316707134246826, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014469081535935402, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014219825156033039, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013741819187998772, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009238814003765583, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19028612971305847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07378249615430832, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06849060952663422, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06556469947099686, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05918467044830322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03409653902053833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.031592488288879395, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04028802365064621, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03718683868646622, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03464261442422867, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030361048877239227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.029121026396751404, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02038855478167534, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017776377499103546, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016412394121289253, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016081975772976875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010232632979750633, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008599404245615005, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008444191887974739, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007865113206207752, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007700666785240173, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005462135653942823, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00557283777743578, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005057182628661394, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0038757447618991137, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07378249615430832, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06555178016424179, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0607106015086174, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05707370117306709, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05136587470769882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.029823213815689087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.027118442580103874, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.037120167165994644, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.034142397344112396, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.030310150235891342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.026620889082551003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025691473856568336, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018632929772138596, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01623375341296196, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01434051338583231, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013858230784535408, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009342196397483349, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00747201032936573, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007225401699542999, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006840784102678299, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00657253572717309, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004989949055016041, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00492122070863843, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004426890052855015, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003269135719165206, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06555178016424179, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22507166862487793, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20970545709133148, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20485655963420868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1850782036781311, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1045415997505188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0987933948636055, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1175910159945488, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10791054368019104, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1060180738568306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0930841937661171, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08885158598423004, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.059478770941495895, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.051355939358472824, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.049822043627500534, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04945946857333183, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.029645731672644615, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025187401100993156, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025065038353204727, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022808372974395752, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02270560711622238, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015153021551668644, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014573989436030388, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014511383138597012, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008714348077774048, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1850782036781311, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16683879494667053, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.134314626455307, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12315312027931213, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10101564228534698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07413134723901749, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06139899790287018, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08897294104099274, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0813567191362381, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07706161588430405, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.053465768694877625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.051017243415117264, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04519263654947281, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0389849878847599, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03572733700275421, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03494046628475189, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02278214506804943, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01897386834025383, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01870916783809662, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015104236081242561, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014649689197540283, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012580415233969688, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012674122117459774, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011606493964791298, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009180879220366478, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16683879494667053, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19124814867973328, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17993561923503876, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17652776837348938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16077330708503723, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08948572725057602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08539583534002304, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09905726462602615, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09157685935497284, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09056517481803894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08068180829286575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07729235291481018, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05040406063199043, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04379072040319443, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04285508394241333, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04263005033135414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02520187385380268, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022162294015288353, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.022099291905760765, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020397430285811424, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02036285027861595, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013348987326025963, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013522380962967873, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0130472956225276, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009214086458086967, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19124814867973328, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24348680675029755, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2293013632297516, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22503986954689026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2050791084766388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11397899687290192, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10882044583559036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12583611905574799, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1165880411863327, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11532743275165558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10283775627613068, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09841132164001465, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06393542885780334, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.055630482733249664, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05444755405187607, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05417225509881973, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03189348801970482, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0277692973613739, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027691777795553207, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025486407801508904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025446567684412003, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01655483804643154, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01635793410241604, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016169730573892593, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010396158322691917, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12583611905574799, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21310503780841827, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1989467293024063, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19308234751224518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17606493830680847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09674243628978729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0917445719242096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11017400026321411, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10183179378509521, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09830685704946518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08800612390041351, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08426356315612793, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.055402591824531555, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04819208011031151, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04599330201745033, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04546627029776573, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027616966515779495, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.023465344682335854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.0232559721916914, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02173484116792679, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021436721086502075, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01435985416173935, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014088032767176628, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013644752092659473, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.008892823942005634, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19308234751224518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07793930172920227, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07251565903425217, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06963802874088287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06296677142381668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.036066435277462006, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03353486210107803, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04224223271012306, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03907972574234009, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03661015257239342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03218851983547211, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030860072001814842, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021354788914322853, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018661972135305405, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017339343205094337, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01701495051383972, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01071401871740818, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009051213040947914, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008899467065930367, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008287820965051651, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008129007183015347, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005710470490157604, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0058076330460608006, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005315849557518959, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0039999582804739475, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07793930172920227, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07231979817152023, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06712383031845093, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06366448104381561, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05734650045633316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03300253301858902, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03031051531434059, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.040003933012485504, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03695466369390488, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.033530060201883316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.029448213055729866, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.028285207226872444, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020054221153259277, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017531171441078186, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015846598893404007, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.015417507849633694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010073117911815643, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00823651161044836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008019648492336273, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00753503805026412, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007300855591893196, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005366922821849585, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005335233174264431, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0048579559661448, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0035568487364798784, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07231979817152023, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23091739416122437, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21543575823307037, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21061593294143677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19046714901924133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10729590058326721, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.101560078561306, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12047547101974487, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11064417660236359, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10882750153541565, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09580250084400177, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09139319509267807, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06099491938948631, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05269787460565567, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.051162756979465485, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05079668015241623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030395841225981712, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0258791446685791, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025762803852558136, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023494843393564224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023409830406308174, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015560157597064972, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015018756501376629, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014930340461432934, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009023362770676613, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19046714901924133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16677579283714294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13888972997665405, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1252085417509079, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10075999051332474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07356621325016022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.060182906687259674, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09610070288181305, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0873139500617981, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07739631086587906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05821503326296806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.052175939083099365, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04895034432411194, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.041984379291534424, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03582092747092247, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.034234724938869476, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024850577116012573, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01957033947110176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01893787272274494, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016795292496681213, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01592319831252098, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014166850596666336, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014140062965452671, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012369781732559204, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010580888949334621, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16677579283714294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19550980627536774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18391582369804382, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18050670623779297, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16437344253063202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09145965427160263, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08727970719337463, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10111585259437561, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09360919892787933, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09256765246391296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08246821165084839, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07892163842916489, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.051395297050476074, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04472438618540764, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04375798627734184, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.043532274663448334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.025675181299448013, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022535691037774086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.022470714524388313, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020719734951853752, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02068432793021202, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013497754000127316, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013614539057016373, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013187996111810207, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009105877950787544, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18391582369804382, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24544616043567657, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23112574219703674, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22683283686637878, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20674720406532288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11495792120695114, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10975005477666855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12704353034496307, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11760227382183075, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11632756888866425, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10371523350477219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09927083551883698, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0645810067653656, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05612775683403015, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.054934754967689514, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05464803799986839, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03223884850740433, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028042752295732498, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027964195236563683, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025736212730407715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025695091113448143, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01681072823703289, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016560254618525505, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016426516696810722, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010576914064586163, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12704353034496307, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21790555119514465, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20356719195842743, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19780460000038147, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1803898960351944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0988866463303566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09393271058797836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11236847937107086, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10376999527215958, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10048698633909225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09007075428962708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08623487502336502, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.056540124118328094, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04914981871843338, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04702258110046387, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.046515487134456635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028270434588193893, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024024033918976784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023825492709875107, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022282598540186882, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021997004747390747, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014868909493088722, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014455236494541168, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01419442892074585, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009211632423102856, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1803898960351944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0924627035856247, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08606231212615967, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0827975794672966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07486208528280258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04291343316435814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03996192663908005, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05014395713806152, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.046225883066654205, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04356492683291435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.038294970989227295, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.036709126085042953, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.025431429967284203, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.022189214825630188, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02068943716585636, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02032095193862915, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012843689881265163, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010905156843364239, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010737044736742973, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010008703917264938, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00983386766165495, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006967391353100538, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007114198990166187, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006516766268759966, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005069661885499954, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0924627035856247, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08045317977666855, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07468412816524506, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07093679159879684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06404954940080643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03672917187213898, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03378714248538017, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04420669004321098, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.040791161358356476, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.037300534546375275, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03272612765431404, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.031464193016290665, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.022210069000720978, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019379232078790665, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.017598725855350494, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.017158938571810722, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011119365692138672, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009099166840314865, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008875031024217606, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008305293507874012, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00806653592735529, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005858753342181444, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005799721460789442, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005307676270604134, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0037759635597467422, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08045317977666855, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2298436313867569, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2142854481935501, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2092570662498474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18906572461128235, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10617700219154358, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10052932798862457, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11926096677780151, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10950669646263123, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10774657130241394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09475730359554291, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09037976711988449, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.060316648334264755, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.052103448659181595, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.050601910799741745, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.050254955887794495, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030056022107601166, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025585956871509552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025473862886428833, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023217925801873207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023130269721150398, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01538220327347517, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014820385724306107, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014750434085726738, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008893974125385284, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18906572461128235, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16630543768405914, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14323300123214722, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1333979070186615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11510534584522247, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07311523705720901, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06354933232069016, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0931766927242279, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08318812400102615, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07651442289352417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06021309271454811, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05694356560707092, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.046838488429784775, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.040998078882694244, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.036529380828142166, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03539842367172241, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02452213689684868, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.021320408210158348, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020909488201141357, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01917857676744461, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01863153837621212, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014849863946437836, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016385376453399658, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013594518415629864, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013714365661144257, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16630543768405914, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19720551371574402, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18555840849876404, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18203739821910858, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16572345793247223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0922989621758461, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08805996179580688, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10204848647117615, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09446916729211807, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09342397004365921, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08317924290895462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07953660935163498, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.051846399903297424, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.045119717717170715, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04414339363574982, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04391331225633621, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.025912294164299965, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022695450112223625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02263041026890278, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02084730938076973, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02081162855029106, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013604843989014626, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013651233166456223, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013291452080011368, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009048025123775005, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18555840849876404, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2473997175693512, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2329573929309845, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22859594225883484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2083045393228531, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11588571965694427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11058870702981949, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12795290350914001, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11855389922857285, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11726920306682587, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10450110584497452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09999784827232361, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06497259438037872, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.056573621928691864, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05536205321550369, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0550759956240654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03244069591164589, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028249062597751617, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028167275711894035, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025915415957570076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02587328664958477, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016872646287083626, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01665741205215454, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016479941084980965, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010600604116916656, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12795290350914001, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21535952389240265, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20126789808273315, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19531577825546265, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17808887362480164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09775639325380325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09276513010263443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11144590377807617, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1029527336359024, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09930883347988129, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08902262896299362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0852055624127388, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05595318600535393, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04873790964484215, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04650748521089554, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04596572741866112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02790517546236515, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0237781573086977, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023557908833026886, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02205217257142067, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021747754886746407, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014507763087749481, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014353901147842407, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013783641159534454, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009168882854282856, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17808887362480164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07013873010873795, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0652531236410141, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06220853328704834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05621100589632988, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03248680382966995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.030001645907759666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.038688864558935165, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03580769896507263, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.032986871898174286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02899164706468582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027850976213812828, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0195637047290802, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017133932560682297, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015678387135267258, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01531971525400877, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009869415313005447, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008294128812849522, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008120767772197723, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007619071751832962, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007434439845383167, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005382632836699486, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005504423752427101, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004961283877491951, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003930389881134033, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07013873010873795, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05983060970902443, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05557916313409805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05196410417556763, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0468655489385128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.027318062260746956, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.024760672822594643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03437427803874016, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.031620848923921585, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02775154821574688, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.024439502507448196, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02371923439204693, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.017240576446056366, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015047552064061165, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013180061243474483, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01269963663071394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008663120679557323, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006915796548128128, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006675030570477247, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006354535464197397, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006091667804867029, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004661086015403271, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00463462807238102, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004106062464416027, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003155677579343319, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05983060970902443, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23286783695220947, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2173624187707901, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21245013177394867, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19212493300437927, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10855495929718018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10273358970880508, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12183409184217453, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11188361793756485, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1100846379995346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09675822407007217, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09232146292924881, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.061731334775686264, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.053302228450775146, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05179881677031517, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05143125355243683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030781716108322144, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02622484415769577, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02610556036233902, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023791009560227394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023699436336755753, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015764331445097923, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015247018076479435, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015119805000722408, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009230384603142738, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19212493300437927, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1379515677690506, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11565450578927994, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10823117941617966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08918503671884537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.060548607259988785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05248982086777687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07317523658275604, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0668710246682167, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06404488533735275, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.047986507415771484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04475829377770424, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.037397172302007675, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03222127631306648, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02939772419631481, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028691723942756653, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0189613476395607, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01596565544605255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01578529179096222, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013681288808584213, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013341959565877914, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010770407505333424, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011102423071861267, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009937060065567493, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008467909879982471, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1379515677690506, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19864997267723083, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18686260282993317, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18327806890010834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16692116856575012, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09301049262285233, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08872625976800919, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10287084430456161, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09518545866012573, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09412378817796707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08379706740379333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08009157329797745, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.052252329885959625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.045465707778930664, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04447752609848976, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04425087943673134, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.026103733107447624, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022860104218125343, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.022793501615524292, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02099987305700779, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.020962882786989212, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01369558647274971, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013742651790380478, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013378334231674671, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009092900902032852, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18686260282993317, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.248134046792984, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23361630737781525, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2292359471321106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20886236429214478, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11623082309961319, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11091328412294388, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12837421894073486, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11892866343259811, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11763133108615875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10481677949428558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10020408034324646, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0652337297797203, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.056738484650850296, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.055520199239254, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.055230170488357544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.032532185316085815, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02827952615916729, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028198661282658577, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025930017232894897, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02588956616818905, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01687515899538994, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01660613901913166, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01647908054292202, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010475738905370235, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12837421894073486, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2178628146648407, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20368774235248566, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19778020679950714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18036909401416779, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09886166453361511, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09390120208263397, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1125093549489975, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10396523028612137, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10041915625333786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09008817374706268, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08624128997325897, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05654235556721687, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04918866604566574, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04698962718248367, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.046461544930934906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028172915801405907, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.023958764970302582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023746877908706665, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02221827395260334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021923039108514786, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01462634652853012, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014354642480611801, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013911825604736805, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009037026204168797, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18036909401416779, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07339394092559814, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06826117634773254, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0652652159333229, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05898205190896988, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03396889194846153, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03145875409245491, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04024611786007881, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.037195801734924316, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03450128436088562, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030305922031402588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.029080478474497795, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020349549129605293, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01778305694460869, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016359476372599602, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016011500731110573, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01022110041230917, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008596031926572323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00843024905771017, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007878039963543415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00770450197160244, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005489978939294815, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0056104701943695545, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005074340384453535, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003932587336748838, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07339394092559814, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06306001543998718, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05861228331923485, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.054823391139507294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.049480509012937546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02886570431292057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.026099052280187607, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03654720261693001, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.033424392342567444, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.029312124475836754, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02583187073469162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02506939321756363, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01834098808467388, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01593061536550522, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013891899026930332, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013367922976613045, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00919495988637209, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007264224346727133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006996814627200365, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006668946240097284, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006381428334861994, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004927007015794516, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004848332144320011, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004330324474722147, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032521423418074846, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06306001543998718, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23555387556552887, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21975450217723846, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2147257775068283, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19409134984016418, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10961809009313583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10365541279315948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12304044514894485, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11307554692029953, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11118469387292862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09765906631946564, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0932074710726738, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06231365352869034, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.053850967437028885, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.052282705903053284, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.051889289170503616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.031076081097126007, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026459388434886932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026340387761592865, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02399151213467121, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023884322494268417, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015914591029286385, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015385694801807404, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015263868495821953, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00930369459092617, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12304044514894485, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16563408076763153, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13931944966316223, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1263076215982437, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10583928972482681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07243594527244568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06122340261936188, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09550002962350845, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08696560561656952, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.077091284096241, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05893655866384506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05619654804468155, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04861285164952278, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04195098206400871, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03535797819495201, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0336303636431694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024633271619677544, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01944269798696041, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018787164241075516, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01703268103301525, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.016079289838671684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013556103222072124, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014249823987483978, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011426400393247604, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010727548971772194, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16563408076763153, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20068027079105377, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18876518309116364, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18516875803470612, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16858476400375366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09397619962692261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08961718529462814, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10385803133249283, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09619568288326263, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09511640667915344, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08464518189430237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08090940862894058, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05277063325047493, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04593491181731224, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.044933393597602844, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04469655081629753, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.026345180347561836, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02305104024708271, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.022985227406024933, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02115679159760475, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02111869305372238, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01376847643405199, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013795855455100536, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013444685377180576, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009050893597304821, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18876518309116364, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2500511109828949, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23535460233688354, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23092061281204224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21040397882461548, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11714204400777817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11178255081176758, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12936580181121826, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11988523602485657, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11855614185333252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10560904443264008, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10095572471618652, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06572927534580231, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.057182636111974716, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05595066398382187, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05566481873393059, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03277837485074997, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02848600596189499, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028406398370862007, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02610837109386921, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026064246892929077, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01697496883571148, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01670863851904869, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016572680324316025, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010503205470740795, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12936580181121826, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22150491178035736, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20752723515033722, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2017005980014801, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18406744301319122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10063929110765457, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0957355797290802, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11414073407649994, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10556620359420776, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10210118442773819, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09182247519493103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08790648728609085, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.057383786886930466, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04992607608437538, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.047799378633499146, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.047285255044698715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028564171865582466, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024312268942594528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02410583756864071, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022563159465789795, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0222761332988739, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01476121973246336, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014462126418948174, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014067726209759712, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.008995589800179005, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18406744301319122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0826871395111084, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07712118327617645, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07407454401254654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06701734662055969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03833306208252907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03571070358157158, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04473922401666641, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.041544340550899506, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.038894228637218475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03427610918879509, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.032823577523231506, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02262384071946144, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.019839096814393997, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01844152621924877, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018095461651682854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011362388730049133, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009643010795116425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009484169073402882, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008848185651004314, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008681521750986576, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006075971759855747, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006209474988281727, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00566404452547431, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004305139649659395, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0826871395111084, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07200592756271362, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06708671152591705, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06337399780750275, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05726239085197449, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03299128636717796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.030190376564860344, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04048644006252289, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03736710548400879, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03347722440958023, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.029527664184570312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.028478607535362244, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020335035398602486, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017776869237422943, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015836425125598907, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.015358194708824158, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010180741548538208, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008220928721129894, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007971261627972126, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007535334676504135, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007275433745235205, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0054122162982821465, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005346501711755991, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004844553302973509, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0035300899762660265, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07200592756271362, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23686642944812775, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22149473428726196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21665558218955994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19629377126693726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1102864146232605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10462816804647446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12369856983423233, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1135452538728714, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11181807518005371, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09865795820951462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09426607191562653, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06266725808382034, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05403916910290718, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05263644456863403, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05228426679968834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.031239621341228485, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026604242622852325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026489952579140663, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0241701677441597, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02409203164279461, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015970377251505852, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01536821760237217, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015350721776485443, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009194796904921532, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12369856983423233, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.20764388144016266, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13790303468704224, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10987978428602219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1048741415143013, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08902842551469803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05862966179847717, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11344488710165024, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10472282767295837, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09677398949861526, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05794898048043251, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.058846741914749146, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.057631053030490875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.050040118396282196, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04299316182732582, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.041175998747348785, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.028840743005275726, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022779790684580803, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02224983461201191, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016977660357952118, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015633556991815567, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015386550687253475, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01564929634332657, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013176124542951584, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010840351693332195, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13790303468704224, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20459333062171936, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19246335327625275, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.188781276345253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17187899351119995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0958782210946083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09145525842905045, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1059955433011055, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09811989217996597, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09702745825052261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08635677397251129, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0826084315776825, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0538976788520813, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04688296094536781, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.045864805579185486, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.045623715966939926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.026910874992609024, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.023584449663758278, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02351965941488743, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02166210301220417, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.021624770015478134, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01411463599652052, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014197582378983498, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013781794346868992, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009418901056051254, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19246335327625275, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25136739015579224, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23660506308078766, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23213724792003632, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2114974707365036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11779122054576874, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11239711940288544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13005943596363068, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12053621560335159, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11921674013137817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10616764426231384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1014765053987503, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06606277823448181, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05749726668000221, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05625820532441139, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05596166476607323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.032942745834589005, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028630945831537247, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0285474956035614, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026230785995721817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026187961921095848, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017044031992554665, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01677005924284458, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01663922145962715, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010510480962693691, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13005943596363068, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22170042991638184, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20756827294826508, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20160812139511108, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18394392728805542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.1006927415728569, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09576163440942764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11474481225013733, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10583848506212234, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10222919285297394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09189537912607193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08797413110733032, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05764539912343025, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.050098177045583725, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.047881968319416046, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.047348957508802414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02880350686609745, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02444584295153618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02423200197517872, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022710075601935387, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02241072617471218, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015115596354007721, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01469290629029274, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014400104060769081, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009327402338385582, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18394392728805542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07600193470716476, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07085923850536346, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06764644384384155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06121424213051796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03534063324332237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03267683833837509, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04209507256746292, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03892538696527481, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.035855624824762344, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03162803128361702, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030419141054153442, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02140560932457447, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018749631941318512, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01714262180030346, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016746604815125465, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010892991907894611, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009215745143592358, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009025187231600285, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008516325615346432, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00831534992903471, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006113582290709019, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006288063246756792, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00566124590113759, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004681411199271679, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07600193470716476, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0675397664308548, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06286526471376419, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05909126251935959, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05328642949461937, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0308661088347435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02813013270497322, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03829182684421539, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.035368625074625015, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03134053945541382, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027640685439109802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02671099081635475, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01918300986289978, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016781063750386238, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014850173145532608, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014354763552546501, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009634007699787617, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007731649558991194, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007475326769053936, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00708808284252882, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006815213244408369, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005120514892041683, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005072943400591612, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004535629879683256, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0033652775455266237, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0675397664308548, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2345181405544281, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21894408762454987, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21404236555099487, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19376631081104279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10910778492689133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10343431681394577, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.122460275888443, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11252707242965698, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11059427261352539, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09745002537965775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09301174432039261, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06199009716510773, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05360715836286545, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0520487055182457, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.051691934466362, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030922219157218933, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026323746889829636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026206430047750473, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023901356384158134, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023807425051927567, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01582062989473343, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015246027149260044, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01518378034234047, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009153272025287151, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19376631081104279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14969578385353088, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11463772505521774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09646621346473694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07923715561628342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06153827905654907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04728526249527931, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08977881819009781, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08109822124242783, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06973803788423538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04866914823651314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.047674838453531265, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.046173177659511566, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04053341597318649, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.032173193991184235, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029882945120334625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024873491376638412, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0204375721514225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019773932173848152, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0185650996863842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01747262477874756, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.016199128702282906, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0175129696726799, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014228909276425838, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.015203170478343964, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14969578385353088, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2046937793493271, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1925242394208908, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18884983658790588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17190510034561157, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09591930359601974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09147627651691437, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10609021782875061, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09818524122238159, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09707631170749664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08638042211532593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08261696994304657, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05393526703119278, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.046905409544706345, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04587976634502411, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04564208164811134, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02694176509976387, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.023573115468025208, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02350671775639057, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.021638283506035805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.021600836887955666, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014147794805467129, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014162464998662472, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01381812896579504, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009358454495668411, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1925242394208908, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2515564560890198, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23673062026500702, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23230186104774475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2115916609764099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11792880296707153, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1124856173992157, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1302233785390854, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1206609308719635, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11933791637420654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10625787824392319, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10156816244125366, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06619244813919067, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.057569634169340134, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05632609874010086, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.056032344698905945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.033003438264131546, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02868393436074257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028598874807357788, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02627849578857422, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026234563440084457, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017096370458602905, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016828302294015884, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016690684482455254, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010594974271953106, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1302233785390854, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22057929635047913, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20673294365406036, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2007058709859848, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.183065265417099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10024438053369522, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09528423845767975, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11459476500749588, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10556142032146454, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10172585397958755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09151051938533783, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08766205608844757, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.057542555034160614, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04997464269399643, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.047668129205703735, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.047118887305259705, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028771191835403442, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024342207238078117, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02411331608891487, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022615956142544746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022302843630313873, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015131264925003052, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014648603275418282, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01438986137509346, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009292135946452618, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.183065265417099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07719065994024277, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0718204453587532, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0687517374753952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.062134332954883575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.035780128091573715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.033169232308864594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.042080752551555634, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03904522955417633, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0363343246281147, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03191862627863884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030538368970155716, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021268034353852272, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01866464875638485, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017216773703694344, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016863225027918816, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010665006935596466, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009015229530632496, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008847962133586407, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00825461745262146, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008079774677753448, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005702495574951172, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0058417594991624355, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0052797854878008366, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0040536364540457726, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07719065994024277, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06438469886779785, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.059969156980514526, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05600471422076225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05059950798749924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.029558300971984863, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02674330770969391, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.037299081683158875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0343264639377594, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.030001217499375343, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.026475703343749046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025538980960845947, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018741199746727943, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01634705625474453, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01423997338861227, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013698352500796318, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009383694268763065, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007439814042299986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007168271578848362, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0068389493972063065, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0065438332967460155, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005022889003157616, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004972381982952356, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004407348111271858, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00333958026021719, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06438469886779785, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24330367147922516, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22694706916809082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22190536558628082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20091751217842102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11347624659538269, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10737498104572296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12733834981918335, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11695709824562073, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11510615050792694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10121053457260132, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09660390019416809, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06452824920415878, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.055712420493364334, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05412942171096802, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.053748439997434616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03216264396905899, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027389001101255417, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02726455219089985, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024826759472489357, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02473440021276474, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01646086759865284, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015865156427025795, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015791745856404305, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009518596343696117, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12733834981918335, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1611364781856537, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1362500637769699, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1244702860713005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10297553986310959, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0717424675822258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06066271662712097, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09066373109817505, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08351478725671768, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07504630088806152, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05565338581800461, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05372168868780136, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04620584845542908, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03990180045366287, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.034579940140247345, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0332157164812088, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.023165537044405937, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.018279438838362694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017713414505124092, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015227516181766987, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014414915814995766, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012534137815237045, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012466474436223507, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01088814064860344, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008617239072918892, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1611364781856537, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20539513230323792, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19317565858364105, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18948028981685638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17245401442050934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0962466150522232, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09180398285388947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10639603435993195, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09853771328926086, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09743206202983856, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08668027073144913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08285406976938248, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.054075539112091064, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.047070156782865524, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04604882746934891, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.045806609094142914, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02700738236308098, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02366909757256508, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02359919808804989, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02172800898551941, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02168966829776764, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014151860028505325, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014230657368898392, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013819769024848938, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009420282207429409, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19317565858364105, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25241759419441223, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23749621212482452, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23303519189357758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21222561597824097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11837010085582733, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11291861534118652, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13071833550930023, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12112125754356384, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11981844902038574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10664303600788116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10193680971860886, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0664290264248848, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.057803183794021606, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.056558024138212204, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05626527592539787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03313757851719856, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02881905995309353, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02873586304485798, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026405755430459976, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026360446587204933, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01717403531074524, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016935143619775772, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016768168658018112, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010698307305574417, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13071833550930023, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21925382316112518, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20524396002292633, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19902272522449493, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18146629631519318, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09961666166782379, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09454550594091415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11390207707881927, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10523918271064758, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10115335136651993, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09088092297315598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08700934797525406, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05732504278421402, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04986317455768585, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.047418344765901566, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04682961106300354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028693359345197678, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024296008050441742, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024049075320363045, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022576460614800453, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022245388478040695, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015151429921388626, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014764603227376938, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014386824332177639, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009521350264549255, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18146629631519318, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06885971873998642, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06402624398469925, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06090264022350311, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05502110347151756, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03189470246434212, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.029352989047765732, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.038207825273275375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03540169075131416, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03239867091178894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02845362387597561, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027289198711514473, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019297605380415916, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.016957104206085205, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01538898702710867, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015003664419054985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009706094861030579, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008123321458697319, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007937818765640259, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007455382030457258, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0072571332566440105, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005246990825980902, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005392875988036394, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004797583445906639, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003816433483734727, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06885971873998642, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0540057048201561, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05019668489694595, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.046405866742134094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.041838787496089935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02475738152861595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02210899069905281, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03221498429775238, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.02947462536394596, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.025154590606689453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.022187383845448494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.021634427830576897, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.016201749444007874, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01406132522970438, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011995217762887478, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011445114389061928, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008185288868844509, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006340489722788334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006052325014024973, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0058499607257544994, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00553639093413949, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00443115783855319, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004373465199023485, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0038199706468731165, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0029977194499224424, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0540057048201561, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2360057681798935, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22025524079799652, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21510227024555206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1946907788515091, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11033264547586441, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10430682450532913, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12406902760267258, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11407104879617691, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11193019151687622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09833782911300659, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0938229039311409, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0628771036863327, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.054355982691049576, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.052682504057884216, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.052256710827350616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03133055567741394, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02667887695133686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02654799446463585, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024196038022637367, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02407323755323887, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01606115698814392, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015549992211163044, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015376324765384197, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009396053850650787, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12406902760267258, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1628851741552353, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1279146671295166, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11242090165615082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09399769455194473, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06965695321559906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.055615223944187164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09352824091911316, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08508703112602234, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07650833576917648, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05436520278453827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05222049355506897, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04777683690190315, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.041507236659526825, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03459306061267853, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03278381749987602, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024386681616306305, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019847726449370384, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019314918667078018, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017281001433730125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.016295485198497772, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014142163097858429, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01535995863378048, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01217496208846569, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012343875132501125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1628851741552353, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2060178965330124, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19375860691070557, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19007821381092072, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1730417013168335, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09659663587808609, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09211602807044983, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10683003067970276, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09888382256031036, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09779156744480133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08696991950273514, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08320693671703339, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.054319847375154495, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04726754128932953, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.046233128756284714, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.045992352068424225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02714373543858528, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.023801831528544426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.023735664784908295, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02185717597603798, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.021816518157720566, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014259389601647854, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01436653546988964, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013926816172897816, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009583011269569397, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19375860691070557, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25310617685317993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23816204071044922, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.233634352684021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21276016533374786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11871838569641113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11326421052217484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13116371631622314, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12151159346103668, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12018221616744995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1069493293762207, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10222091525793076, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0666789561510086, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05800451338291168, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.056750085204839706, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.056456271559000015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.033251211047172546, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028949523344635963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02886536717414856, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026526805013418198, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026481743901968002, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017254704609513283, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017062030732631683, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016842573881149292, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010844417847692966, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13116371631622314, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2199457734823227, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2060893177986145, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19974254071712494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.182180255651474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09997100383043289, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09486983716487885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11461997777223587, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10577093809843063, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10144913196563721, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09123753756284714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08736725896596909, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.057532306760549545, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.050080832093954086, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0475790835916996, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04698025807738304, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02869364246726036, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02437947690486908, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024122942239046097, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022663649171590805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022324688732624054, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014989901334047318, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014828339219093323, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014189091511070728, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009557455778121948, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.182180255651474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07327460497617722, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06809189170598984, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06483667343854904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05860823392868042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03405392915010452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03136252984404564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.041011348366737366, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03769403696060181, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03460612893104553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030367977917194366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02926568314433098, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020849090069532394, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018108857795596123, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01648571342229843, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016086602583527565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010561486706137657, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008750833570957184, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008560534566640854, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00804054457694292, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007836645469069481, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0057610319927334785, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005837175063788891, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005295990966260433, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004181930795311928, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07327460497617722, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05670996010303497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05268007516860962, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04855741932988167, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04379484802484512, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.025932330638170242, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.023135030642151833, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.033748045563697815, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.031023431569337845, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.026351090520620346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.023222602903842926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.022575005888938904, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01691901683807373, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.014737670309841633, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.012522093951702118, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01194529328495264, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008503822609782219, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006603608839213848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006311346311122179, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006078131031244993, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005758644547313452, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004583076573908329, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004541229922324419, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003937116824090481, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003105571959167719, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05670996010303497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2322433441877365, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21633808314800262, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.211204394698143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19090117514133453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10822415351867676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10218771547079086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12168392539024353, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11183639615774155, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10974497348070145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09625015407800674, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09169907122850418, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06159735098481178, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05325116217136383, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.051604144275188446, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05120998993515968, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03069736808538437, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02609093114733696, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025957828387618065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023603789508342743, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023487543687224388, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015692221000790596, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015124510042369366, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015028832480311394, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009033439680933952, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19090117514133453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1624942421913147, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12803469598293304, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11158530414104462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09311151504516602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07081267237663269, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.056165698915719986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09388251602649689, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0857812836766243, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07528533041477203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0526302196085453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05198288708925247, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.047918058931827545, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04135333374142647, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03473601117730141, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03300091624259949, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02425527200102806, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019371192902326584, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01870555989444256, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01623588055372238, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01521277241408825, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013712147250771523, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01447815727442503, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011743237264454365, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011177738197147846, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1624942421913147, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20684273540973663, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19450612366199493, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19076763093471527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17359140515327454, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09704176336526871, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09251732379198074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10730589926242828, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09933993220329285, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.098233662545681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08731647580862045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0834539458155632, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.054514091461896896, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.047477565705776215, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.046438101679086685, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.046193677932024, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027228962630033493, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.023889122530817986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02381964586675167, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.021919405087828636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02188180759549141, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014253432862460613, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014388361014425755, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013921019621193409, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.00955625157803297, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19076763093471527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2549036741256714, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23978523910045624, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23524805903434753, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21414655447006226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11961356550455093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11407569795846939, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13207271695137024, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12241970747709274, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12106577306985855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.107671819627285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10285115987062454, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0670916959643364, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.058415766805410385, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.057152844965457916, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.056855812668800354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03346680849790573, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02911335602402687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029030438512563705, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026659924536943436, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026612883433699608, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01731862872838974, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017100531607866287, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016903989017009735, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01078135147690773, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13207271695137024, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21887272596359253, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2050735056400299, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19873113930225372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18125087022781372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09946641325950623, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09439216554164886, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11379622668027878, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10527870059013367, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10098867118358612, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09082727879285812, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08692074567079544, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05724009871482849, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04990067705512047, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04734908416867256, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04674433544278145, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028594933450222015, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024272538721561432, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.0240047425031662, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02257128432393074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022219566628336906, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014982483349740505, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01478519756346941, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01418590359389782, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009527858346700668, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18125087022781372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.053554121404886246, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.049510370939970016, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.046040285378694534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.041488710790872574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.02468767762184143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.022153398022055626, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.031106680631637573, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.028859255835413933, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.02513962984085083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02195730432868004, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02111044153571129, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.015659239143133163, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.013784981332719326, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01194791030138731, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.011475353501737118, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00790963415056467, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.006370247807353735, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.006137790624052286, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00584780378267169, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.005592772271484137, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004347712732851505, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004407195374369621, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.003837837837636471, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0031400606967508793, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.053554121404886246, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.03233528509736061, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.029963508248329163, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.02541959099471569, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.022986173629760742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.01461187656968832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.011889416724443436, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.022101949900388718, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.02020951360464096, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.014867802150547504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.013208451680839062, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.013095930218696594, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.011007238179445267, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.009630276821553707, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.007136540021747351, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.006432687398046255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.005581979174166918, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.003929379396140575, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.003568297019228339, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0036776126362383366, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.003287923987954855, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0031729615293443203, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003068353049457073, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.002526064869016409, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0021913452073931694, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.03233528509736061, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23941795527935028, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22255878150463104, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2171211987733841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19589613378047943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11186233907938004, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10535070300102234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12617798149585724, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11581576615571976, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11361576616764069, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09920629113912582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09449025988578796, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06398387998342514, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05522235482931137, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.053399793803691864, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0529763400554657, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03189830109477043, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027001840993762016, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026865597814321518, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024368032813072205, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024235254153609276, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0163172148168087, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015680933371186256, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015561687760055065, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009359178133308887, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12617798149585724, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.15337195992469788, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12931522727012634, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11390220373868942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10504937916994095, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06833826005458832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05362703651189804, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09317111223936081, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08541172742843628, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07179195433855057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05789482966065407, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05675768107175827, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04745236784219742, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04096261411905289, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.033335570245981216, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.031274374574422836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0239418875426054, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01819327473640442, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017284860834479332, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016283854842185974, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015091951936483383, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01320045068860054, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013405500911176205, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010899948887526989, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009756684303283691, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.15337195992469788, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20807422697544098, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1955796629190445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19182606041431427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1745651662349701, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09764878451824188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09308339655399323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10804817825555801, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09997440874576569, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.0988483801484108, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08785166591405869, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0840575248003006, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05490142107009888, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04780369624495506, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04675457999110222, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.046510133892297745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027464747428894043, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02410302683711052, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024034352973103523, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02212568186223507, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022082308307290077, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01446276530623436, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01459517702460289, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014123683795332909, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009787491522729397, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19182606041431427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2562157213687897, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2409622073173523, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23639820516109467, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21517935395240784, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12027575075626373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11468518525362015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13282494246959686, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12310288846492767, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12175827473402023, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10825599730014801, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10342517495155334, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06751005351543427, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.058749206364154816, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05747256055474281, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05717381089925766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03367217630147934, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029288705438375473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029204465448856354, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026815040037035942, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026769183576107025, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017436135560274124, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017218278720974922, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01701805554330349, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010875155217945576, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13282494246959686, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2219759225845337, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2080119401216507, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20141303539276123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1836649775505066, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10086199641227722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09566192328929901, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11616842448711395, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10702840983867645, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10237714648246765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09212420880794525, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08830252289772034, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.058336272835731506, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05070710927248001, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.048038456588983536, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.047392964363098145, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02919125184416771, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02466941997408867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024388672783970833, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02295665629208088, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022592587396502495, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015374995768070221, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01510397344827652, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01451188512146473, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009828021749854088, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1836649775505066, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06908351182937622, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06409836560487747, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.061087191104888916, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05515548214316368, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0320470854640007, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02954743057489395, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03816332668066025, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03534487262368202, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03256933018565178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02850404381752014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02728421613574028, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019284401088953018, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0169123075902462, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015462325885891914, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015101620927453041, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009706126525998116, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008169157430529594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007999411784112453, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007482095155864954, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007301074918359518, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005259043537080288, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005408701952546835, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004833413753658533, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0038456202019006014, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06908351182937622, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04919419810175896, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04565390571951866, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04189947620034218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.037754692137241364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0224577896296978, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.019948257133364677, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.029715584591031075, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.027037084102630615, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.022819053381681442, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.020080100744962692, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.019558757543563843, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.014880253002047539, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.012851333245635033, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0108519671484828, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.010318818502128124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007502795197069645, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.005726404953747988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00545099750161171, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005272085778415203, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.004973651375621557, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0040588416159152985, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003960547503083944, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003453704295679927, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002687514293938875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04919419810175896, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24828574061393738, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2313467264175415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2259254902601242, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20403076708316803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11588086187839508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10951431095600128, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12995314598083496, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11956235021352768, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1175844669342041, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10301648080348969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09809406101703644, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06579651683568954, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0569918118417263, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.055321983993053436, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05490965396165848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.032792285084724426, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02796429954469204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027838628739118576, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.025296900421380997, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025179116055369377, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016776204109191895, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01620415225625038, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016118831932544708, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009674584493041039, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12995314598083496, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.15951350331306458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13267384469509125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1167985126376152, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10287123918533325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06788327544927597, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.055523838847875595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09491299837827682, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08733338117599487, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07500994205474854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05829484760761261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.056482840329408646, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.048328012228012085, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.041771307587623596, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03303058445453644, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.030588263645768166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024262502789497375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017794856801629066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016913730651140213, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016002483665943146, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014567380771040916, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013095485046505928, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013096894137561321, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010374611243605614, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009086077101528645, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.15951350331306458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2092391401529312, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19664698839187622, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1928449124097824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1754050999879837, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09823325276374817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09361603111028671, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10869753360748291, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1005978211760521, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09946443885564804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08833670616149902, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0844450518488884, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.055274128913879395, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.048098355531692505, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.047041554003953934, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.046789247542619705, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027627112343907356, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02421887218952179, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02414785884320736, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022219369187951088, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02218015491962433, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014533879235386848, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014628211967647076, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014191728085279465, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009758565574884415, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1928449124097824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25783538818359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2425047606229782, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23786616325378418, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2164159119129181, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12105722725391388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11538858711719513, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13374152779579163, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12391239404678345, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12254166603088379, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10891548544168472, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10405236482620239, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06800661981105804, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.059141211211681366, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0578535795211792, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05754634365439415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03391934186220169, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02946564368903637, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02938259206712246, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026963060721755028, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026918230578303337, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017585506662726402, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01729249395430088, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017163865268230438, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010882532224059105, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13374152779579163, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22447195649147034, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21036465466022491, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2037968933582306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18588699400424957, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10194005817174911, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09676431864500046, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11687987297773361, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10801870375871658, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10351729393005371, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09315058588981628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08919951319694519, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.058766331523656845, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0510978177189827, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.048476092517375946, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.047841936349868774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02928830310702324, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02474142052233219, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02447110414505005, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023003598675131798, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02264174073934555, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015237671323120594, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014925304800271988, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01438391488045454, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009418022818863392, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18588699400424957, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07379516214132309, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06845912337303162, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06530772149562836, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05891617387533188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03423900529742241, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03157183900475502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.040714018046855927, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.037711597979068756, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03480333089828491, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030451782047748566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.029140153899788857, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02058635838329792, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01806275174021721, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01654740795493126, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016171708703041077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010355733335018158, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008763543330132961, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008586719632148743, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008029205724596977, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007841888815164566, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005606741178780794, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005827935878187418, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005160434637218714, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00418238015845418, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07379516214132309, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05383812636137009, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.049936916679143906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04620755463838577, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04166148602962494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.024719173088669777, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.022115109488368034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03217904642224312, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.029335878789424896, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.025135604664683342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02208910509943962, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02149994857609272, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0161572452634573, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.013988569378852844, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011961154639720917, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011437010951340199, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00810934230685234, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006333672441542149, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0060651167295873165, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005823868326842785, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005532254930585623, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004394303075969219, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00436599925160408, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0037857985589653254, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00302226934581995, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05383812636137009, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24772052466869354, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23104453086853027, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22556568682193756, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20388992130756378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11601719260215759, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10959118604660034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13003411889076233, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11977775394916534, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1177079826593399, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10313813388347626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09824585914611816, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06590167433023453, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05707734078168869, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.055370453745126724, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05500602722167969, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03288113698363304, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0279864352196455, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02785678394138813, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.025306610390543938, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02519059367477894, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016803201287984848, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016186002641916275, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016139322891831398, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009622919373214245, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13003411889076233, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14228442311286926, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11942610144615173, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11208216845989227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09781382232904434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06305265426635742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05432779714465141, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07466263324022293, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06864254176616669, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06606311351060867, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05111530050635338, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.048142556101083755, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03789401054382324, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03274204209446907, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.030253488570451736, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02964061312377453, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018971385434269905, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015823962166905403, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015661034733057022, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01350833848118782, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013193515129387379, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010147814638912678, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010211254470050335, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009354052133858204, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007052883971482515, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14228442311286926, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2103077471256256, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19759175181388855, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19379715621471405, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1761770248413086, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09878181666135788, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0941266119480133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10935328900814056, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10116878151893616, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10001157224178314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08881253749132156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08490267395973206, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.055659614503383636, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.048384394496679306, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04731902852654457, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04706387221813202, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0278158001601696, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024397922679781914, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02432786487042904, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02238503471016884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022347373887896538, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014673367142677307, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014786888845264912, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014333522878587246, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009934274479746819, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19379715621471405, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.258827805519104, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2433548867702484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23871292173862457, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21718379855155945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12156903743743896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11586213856935501, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13426028192043304, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12443044781684875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12306513637304306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10931555926799774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10440441220998764, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06824842840433121, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05938189849257469, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.058080028742551804, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.057774510234594345, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03403782844543457, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02957628294825554, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02949078567326069, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02705710008740425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02701280079782009, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017622020095586777, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017354372888803482, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017197972163558006, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01090854313224554, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13426028192043304, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2260512411594391, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.211924210190773, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2051449418067932, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18715761601924896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10276086628437042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09747466444969177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11838842928409576, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1091541200876236, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10431316494941711, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09393425285816193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08997691422700882, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05934170261025429, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05171097442507744, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04897616058588028, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04831136763095856, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02957056649029255, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02520797960460186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024915175512433052, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023481644690036774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023102421313524246, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015367059037089348, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015515860170125961, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01445904839783907, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010187122970819473, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18715761601924896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0702492892742157, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06525015830993652, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06182084232568741, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05580395460128784, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.032597798854112625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02985755354166031, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03981497138738632, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03652322292327881, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03311585634946823, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.029060175642371178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027977274730801582, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020170336589217186, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017545728012919426, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015795432031154633, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015361514873802662, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010226583108305931, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008423181250691414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008210267871618271, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0077451663091778755, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00751930195838213, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005626589525490999, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005692995153367519, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005123418755829334, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004099657759070396, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0702492892742157, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05007247254252434, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.046452637761831284, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.042707331478595734, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03842887282371521, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02298693172633648, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.020421333611011505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.030313093215227127, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.02763059176504612, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02336272969841957, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02056261897087097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.01996038667857647, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01524101011455059, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01318507082760334, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011125599965453148, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.010576257482171059, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00769102293998003, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.005880721379071474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00560699263587594, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005413518752902746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005118229426443577, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004172096960246563, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0040772599168121815, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003571164794266224, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002794103929772973, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05007247254252434, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.235503152012825, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21928797662258148, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21411801874637604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19331584870815277, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11003884673118591, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1038968414068222, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12364665418863297, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11377759277820587, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11166682094335556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09767536073923111, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09306055307388306, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06263251602649689, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.054272666573524475, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.052554018795490265, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.052145738154649734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.031240632757544518, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02667143940925598, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0265291016548872, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0241079181432724, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02399880811572075, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01603427529335022, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015597296878695488, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01538415253162384, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009533791802823544, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19331584870815277, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1587098240852356, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1366402804851532, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1159185916185379, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10112880915403366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07066711038351059, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.055850204080343246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10501451045274734, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09599606692790985, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07484758645296097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06090965494513512, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05992034077644348, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05438308045268059, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.047267571091651917, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03618951141834259, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03313437104225159, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0284532830119133, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02237679250538349, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02108842507004738, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02087416872382164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01932629756629467, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.016916519030928612, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.018981914967298508, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013782408088445663, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.015921976417303085, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1587098240852356, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21039122343063354, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19766473770141602, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19383777678012848, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17628945410251617, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09884887933731079, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09418029338121414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10948240756988525, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10121544450521469, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10007905960083008, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08885624259710312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08494111895561218, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05571303516626358, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04842634126543999, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.047363199293613434, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.047107014805078506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02788330242037773, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0244310162961483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024361038580536842, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022416677325963974, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022377075627446175, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014761828817427158, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014824047684669495, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014418135397136211, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.00997859239578247, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17628945410251617, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2577201724052429, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24227121472358704, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23767304420471191, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21617485582828522, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12105052173137665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11537545919418335, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13374611735343933, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12390568107366562, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12256380915641785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10885138064622879, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10394403338432312, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0679963231086731, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.059144314378499985, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05785397067666054, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.057546623051166534, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.033912889659404755, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02945154532790184, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02936818264424801, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02693871408700943, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026891928166151047, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017566809430718422, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017271611839532852, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017141882330179214, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010845722630620003, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13374611735343933, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22396469116210938, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20978142321109772, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20303919911384583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18510770797729492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10181533545255661, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09649692475795746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11703946441411972, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10820046067237854, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10339757800102234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09300124645233154, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08902312070131302, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0588456466794014, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0513547845184803, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04855361580848694, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04788227006793022, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029422186315059662, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025043264031410217, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024750549346208572, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023322729393839836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022944653406739235, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01542778592556715, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015498479828238487, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014532293193042278, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010273146443068981, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18510770797729492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07598679512739182, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07061873376369476, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06743276864290237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06094659864902496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03529896214604378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0326329730451107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04181749373674393, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03877902403473854, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03585280850529671, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03145481273531914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030129220336675644, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021163392812013626, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018550194799900055, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017028188332915306, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016661854460835457, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010643329471349716, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008992068469524384, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008816658519208431, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008244045078754425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008059733547270298, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005753010511398315, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005930696614086628, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005301680415868759, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004217042587697506, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07598679512739182, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06052903085947037, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.056151896715164185, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05232585221529007, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04718904569745064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.027694208547472954, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.024929286912083626, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.035344887524843216, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03236674144864082, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.028123756870627403, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.024733247235417366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.023943740874528885, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.017704380676150322, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015396314673125744, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013323626480996609, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.012794947251677513, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008887356147170067, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006975878495723009, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006697025150060654, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0063923257403075695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006093897856771946, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004758617375046015, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004677896853536367, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004122559912502766, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003127789357677102, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06052903085947037, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2478015273809433, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23107999563217163, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22572903335094452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20410485565662384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11568816006183624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10939061641693115, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12961065769195557, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11933133006095886, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11738541722297668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10301228612661362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09819736331701279, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06567014753818512, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05686023458838463, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.055200546979904175, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.054809994995594025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.032739344984292984, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02789454720914364, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027761097997426987, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02524527721107006, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025148779153823853, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016726579517126083, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016108369454741478, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016068676486611366, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009579562582075596, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12961065769195557, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18807801604270935, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13063618540763855, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10515714436769485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09622161090373993, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07754216343164444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05179799348115921, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10709158331155777, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0979444608092308, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08745972812175751, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05594008415937424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05655404552817345, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0546627901494503, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04687217250466347, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03771815821528435, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.035237014293670654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027429064735770226, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02041439525783062, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01967260055243969, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016509374603629112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014840634539723396, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014888618141412735, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014987760223448277, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011979855597019196, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01064060628414154, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18807801604270935, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20887354016304016, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19619004428386688, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1923593282699585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1749275177717209, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09819246828556061, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09353357553482056, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10862839221954346, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10057186335325241, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09942423552274704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08822514861822128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08428750187158585, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.055305469781160355, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04811060428619385, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04703653231263161, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04678591340780258, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02763347327709198, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024265017360448837, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02419278770685196, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02225351519882679, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022215910255908966, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01456491556018591, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01471780426800251, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014219913631677628, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009904790669679642, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1923593282699585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25697892904281616, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24154461920261383, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23689626157283783, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21545249223709106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12073679268360138, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11505775153636932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1334335058927536, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12362446635961533, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12225332111120224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10856088995933533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10368721932172775, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06783192604780197, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05904082953929901, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05774267017841339, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.057435303926467896, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03388093784451485, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029472801834344864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02938275784254074, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026962200179696083, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026916271075606346, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017647113651037216, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017389213666319847, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017222171649336815, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011083749122917652, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1334335058927536, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2228456437587738, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2085667848587036, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20163516700267792, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18386805057525635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10118258744478226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09578967839479446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1166384220123291, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10785996168851852, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10283613950014114, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09241894632577896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08851060271263123, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05865126848220825, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0510828010737896, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.048165448009967804, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04745974764227867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029305480420589447, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024674419313669205, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024366632103919983, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02295086532831192, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022546488791704178, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015350441448390484, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01507001556456089, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01442761905491352, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009661625139415264, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18386805057525635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0629103034734726, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.058223262429237366, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.055100150406360626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04965438321232796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.029135920107364655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.026644473895430565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03531118482351303, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0327024944126606, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.02964736893773079, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.025866298004984856, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02478327415883541, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.017839372158050537, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.015624459832906723, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014050889760255814, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.013662458397448063, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00897256750613451, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007405094336718321, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007215210236608982, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006765913683921099, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006558221764862537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004845572169870138, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00492126727476716, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0043816533870995045, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003437899751588702, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0629103034734726, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04029234126210213, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.03719153255224228, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.033555541187524796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03019816055893898, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.01831701584160328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0159732848405838, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.025182396173477173, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.02282469905912876, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.018651174381375313, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.016386130824685097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.01610146090388298, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.012636822648346424, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.010840828530490398, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.00889842864125967, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.008358667604625225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.006391278002411127, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00473566772416234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.004456366878002882, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.004368758760392666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.004061745945364237, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0034924130886793137, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0033798418007791042, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0029125551227480173, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0023217271082103252, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04029234126210213, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24607619643211365, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2288169413805008, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22330302000045776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2013448029756546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11509508639574051, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10860170423984528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12903441488742828, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11886879801750183, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11685868352651596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10198189318180084, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09690698981285095, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0654059424996376, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.056719668209552765, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05498703569173813, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05458072945475578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.032629139721393585, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027845721691846848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027711419388651848, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02510417066514492, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024996232241392136, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01671348139643669, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01619414985179901, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01605704426765442, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009787528775632381, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12903441488742828, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16993771493434906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14318245649337769, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1337423324584961, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11465957015752792, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07626470178365707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06503107398748398, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09159085154533386, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08380649238824844, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07998770475387573, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.060878120362758636, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05719470977783203, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04665491357445717, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0400545597076416, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03665876388549805, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03581859543919563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02334575727581978, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01911519281566143, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018864555284380913, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016080079600214958, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015644360333681107, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012399137951433659, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012320471927523613, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011327890679240227, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008358634077012539, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16993771493434906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2105429470539093, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19776476919651031, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19390679895877838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17623253166675568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09897977858781815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09427940845489502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10949911922216415, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10136792808771133, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10021258890628815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08892577141523361, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0849565714597702, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.055718425661325455, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.048492107540369034, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04741176962852478, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.047155849635601044, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0278526209294796, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024442363530397415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024370945990085602, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02241247147321701, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022373290732502937, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014668076299130917, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01480405405163765, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014318862929940224, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009932808578014374, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17623253166675568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25653156638145447, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24104447662830353, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2364228516817093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2149820774793625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12058156728744507, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1148669570684433, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13331615924835205, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12342564761638641, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12204857915639877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10836021602153778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10349703580141068, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06785086542367935, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.058955442160367966, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05766264349222183, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05735465884208679, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03386811539530754, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029437024146318436, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02935105189681053, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026930980384349823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026883702725172043, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017677253112196922, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01738778129220009, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017255248501896858, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011098741553723812, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13331615924835205, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2231239229440689, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20851324498653412, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.201497420668602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18371902406215668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10119825601577759, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09578901529312134, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11683129519224167, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1079631969332695, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10295851528644562, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09238557517528534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08850692957639694, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05873960256576538, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05107109993696213, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04815225303173065, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04744405671954155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02926279604434967, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024607859551906586, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02430533803999424, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02286674827337265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022459905594587326, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015214650891721249, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014945808798074722, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014265441335737705, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009454061277210712, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18371902406215668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07024163752794266, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0652032196521759, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06186100095510483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05580286309123039, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.032570142298936844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.029872583225369453, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03930553048849106, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.036403194069862366, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03309306502342224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.028988514095544815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027791159227490425, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019866906106472015, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017413485795259476, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015721553936600685, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015308011323213577, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00998610444366932, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008301790803670883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008096839301288128, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007609616033732891, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00738955894485116, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005388804245740175, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005530879367142916, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004885038360953331, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0039018664974719286, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07024163752794266, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0500776544213295, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.046499066054821014, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04241180419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.038162101060152054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02285756915807724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02012544311583042, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.030725911259651184, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.028059832751750946, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.023224828764796257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.020486893132328987, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.01997501216828823, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.015416854992508888, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.013339096680283546, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011050605215132236, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.010443716309964657, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007745051756501198, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0058339983224868774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.005515480879694223, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005370727274566889, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005027285311371088, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004156988114118576, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004075844772160053, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0034473359119147062, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0027465003076940775, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0500776544213295, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2515832483768463, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23444287478923798, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22901779413223267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20682963728904724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11765378713607788, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11118483543395996, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.131818026304245, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12149343639612198, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11933590471744537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10455554723739624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09946125000715256, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06679962575435638, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05790626257658005, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05614965409040451, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.055755097419023514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03330102190375328, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02836555428802967, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02822992205619812, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.025641944259405136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025519274175167084, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.017026236280798912, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016416329890489578, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016359860077500343, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009762153029441833, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.131818026304245, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1695500612258911, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11515742540359497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09162016957998276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09065769612789154, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07014114409685135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04281453415751457, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09494908154010773, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08747610449790955, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07929722964763641, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0517096109688282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.052055615931749344, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04840155690908432, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04192064702510834, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03410784900188446, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03200838714838028, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02430487610399723, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.018328800797462463, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017754163593053818, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014927727170288563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013521619141101837, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013055309653282166, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013252210803329945, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010580343194305897, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00929716695100069, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1695500612258911, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2105548232793808, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19773440062999725, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19381137192249298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17618292570114136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09902405738830566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09428799897432327, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10956543684005737, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10141266882419586, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10026060044765472, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08892205357551575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08491915464401245, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05577532947063446, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04851996898651123, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04743880778551102, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.047184064984321594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027889274060726166, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02445833943784237, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024386610835790634, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02242075279355049, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022379927337169647, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014705904759466648, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014817945659160614, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014356984756886959, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009947169572114944, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19381137192249298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25494635105133057, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23951959609985352, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2349204421043396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2135537713766098, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11981266736984253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11414080858230591, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13241179287433624, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12265434861183167, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12132135033607483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10766282677650452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10273650288581848, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06731098890304565, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.058562375605106354, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.057273343205451965, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.056964434683322906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03357395902276039, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029176246374845505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02909225970506668, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026669111102819443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026626694947481155, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017393959686160088, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017138034105300903, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016970213502645493, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01079621259123087, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13241179287433624, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20738694071769714, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19359874725341797, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18691357970237732, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1704065054655075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09406247735023499, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08888460695743561, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10873765498399734, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1005803793668747, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09575428813695908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08582916855812073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08221085369586945, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.054661501199007034, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.047689467668533325, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04483817517757416, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0441519096493721, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027291705831885338, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0230710506439209, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02277359552681446, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02146228216588497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021063972264528275, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014271403662860394, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01425595860928297, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013362019322812557, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009312410838901997, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19359874725341797, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.059447161853313446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.055326879024505615, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05098532885313034, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.046112414449453354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.027685729786753654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02466285042464733, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03601144254207611, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03302256762981415, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.02811550907790661, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.024738943204283714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02404877543449402, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01830894686281681, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01590963453054428, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.013558453880250454, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.012944936752319336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00934857502579689, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007453261874616146, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0071441372856497765, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006912002339959145, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006578088738024235, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0053043426014482975, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00543964421376586, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004634440410882235, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004115532152354717, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.059447161853313446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.03946234658360481, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.036597803235054016, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.032386984676122665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.029094429686665535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.01795956678688526, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.015326332300901413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.02507193200290203, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.023201201111078262, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.018272966146469116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.016045868396759033, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.015653187409043312, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.012598729692399502, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.011008725501596928, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.008719734847545624, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.00807289220392704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.006387363653630018, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.004657907877117395, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.004333370830863714, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.004296023398637772, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0039366609416902065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.003551891539245844, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003404576564207673, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0029133602511137724, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0023417414631694555, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.03946234658360481, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2473084181547165, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22946970164775848, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22356092929840088, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20071136951446533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1156061515212059, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10875274986028671, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13057315349578857, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11985915154218674, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11743468046188354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10196219384670258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09686990082263947, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06625179201364517, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0571945495903492, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.055250879377126694, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05480922386050224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03306799754500389, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028044790029525757, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027887098491191864, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02523723989725113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025103604421019554, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016971971839666367, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016462109982967377, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016175296157598495, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010083537548780441, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13057315349578857, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0725359320640564, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.05793321877717972, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.04825033247470856, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04076523333787918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.030958713963627815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.023659221827983856, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04711884632706642, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04169687628746033, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03393586352467537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.024997767060995102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.025258978828787804, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02460475265979767, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.020642993971705437, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015989335253834724, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014723986387252808, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013218765147030354, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010010740719735622, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009554223157465458, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009130576625466347, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008481023833155632, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008673137053847313, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008564612828195095, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0075754281133413315, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00730340089648962, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0725359320640564, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21115924417972565, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19826659560203552, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19439783692359924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17662087082862854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09931348264217377, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09455014765262604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10984469950199127, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10171034187078476, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10055707395076752, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08916912972927094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08514400571584702, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.055886704474687576, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04865168407559395, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.047564733773469925, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04730759561061859, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02793053165078163, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024499673396348953, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024428684264421463, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02245291694998741, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022411445155739784, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014663885347545147, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01481294259428978, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014315318316221237, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009902441874146461, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17662087082862854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25497955083847046, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23957017064094543, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23489248752593994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21354499459266663, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11988582462072372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11420300602912903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13247482478618622, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12275005131959915, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12139149755239487, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10770256817340851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10279539972543716, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06738275289535522, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.058618076145648956, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.057323239743709564, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.057018477469682693, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03362107276916504, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02923981472849846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02915494702756405, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026733523234725, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026684967800974846, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017450762912631035, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017226891592144966, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017030078917741776, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010934850201010704, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13247482478618622, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21868017315864563, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.203886017203331, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19661487638950348, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17910631000995636, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09908584505319595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09349822252988815, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11525600403547287, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10628422349691391, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10094401985406876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09033006429672241, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08652398735284805, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0578998401761055, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.050327979028224945, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04721663147211075, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04646177217364311, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02898729406297207, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024269500747323036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02394202910363674, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02255486138164997, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022115975618362427, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015353763476014137, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014974026009440422, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014372169971466064, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009717310778796673, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17910631000995636, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08767589181661606, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08158152550458908, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07856777310371399, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07094256579875946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04085297882556915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0380902960896492, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.047405093908309937, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04387202858924866, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04148950055241585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03636234998703003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03469755873084068, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.024005893617868423, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.020974507555365562, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0196464192122221, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.019325189292430878, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012027889490127563, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010238108225166798, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010092142038047314, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009347205981612206, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009197910316288471, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006393093150109053, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006515382323414087, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00598389096558094, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00447422917932272, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08767589181661606, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07629062980413437, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0708208680152893, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06736684590578079, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06063941493630409, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.035104360431432724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.032297905534505844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04259426146745682, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03896147757768631, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.035676829516887665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.031216051429510117, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03000812977552414, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.021434767171740532, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.018543750047683716, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016870059072971344, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016436269506812096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010734685696661472, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008727576583623886, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008519692346453667, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007955928333103657, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007730455603450537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005669026169925928, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005581301636993885, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005126133561134338, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0036696025636047125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07629062980413437, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2587950527667999, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24106687307357788, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23555369675159454, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2126518040895462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12086160480976105, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11431798338890076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13483165204524994, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12450426816940308, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1226336658000946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1073908656835556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10211089998483658, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06838906556367874, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05936574935913086, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.057705868035554886, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05733883008360863, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03408977761864662, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02931641973555088, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.029184075072407722, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.026542965322732925, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0264544989913702, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.017507487908005714, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01714378036558628, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016896221786737442, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010553540661931038, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13483165204524994, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19045549631118774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13050392270088196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10868638008832932, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10405613481998444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08156684786081314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05363166704773903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10231859982013702, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09380390495061874, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08868897706270218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05634111166000366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05624419450759888, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0519338995218277, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0447748601436615, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.039288606494665146, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03790055587887764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026042118668556213, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02069719508290291, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020373882725834846, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01593773625791073, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014965839684009552, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013995293527841568, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013941030949354172, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012273947708308697, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009609682485461235, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19045549631118774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2089076191186905, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1961584836244583, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19225981831550598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17476233839988708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09824787825345993, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09356427192687988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10869353264570236, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10063926130533218, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09949284046888351, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0882156491279602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08423469960689545, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05534786358475685, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.048157479614019394, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0470866933465004, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0468343086540699, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027666902169585228, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0242929644882679, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024226147681474686, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022268489003181458, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022228475660085678, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014576917514204979, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01474611833691597, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014229215681552887, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009927737526595592, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19225981831550598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25160959362983704, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23635555803775787, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23178841173648834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21070441603660583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1183021068572998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11270316690206528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13091084361076355, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12111630290746689, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11977826058864594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10625427961349487, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10142163932323456, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06659485399723053, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05787181854248047, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.056594956666231155, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.056292980909347534, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03331107273697853, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028912346810102463, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028830379247665405, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026437127962708473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026392709463834763, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017479484900832176, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01710480824112892, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01706736348569393, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010956735350191593, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13091084361076355, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2225293219089508, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20675814151763916, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1988658607006073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18115219473838806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10068272799253464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09469228237867355, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11776679009199142, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10853780061006546, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10275217890739441, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09160755574703217, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08778774738311768, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05914977192878723, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0514080673456192, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.048010971397161484, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04718688502907753, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029503824189305305, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024724474176764488, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024375060573220253, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022947754710912704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022468237206339836, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015461865812540054, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015362142585217953, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014362595975399017, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010039852000772953, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18115219473838806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09883175790309906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0920923575758934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0889747366309166, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08052904158830643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.046110209077596664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.043151162564754486, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.053655579686164856, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.049120593816041946, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04679224640130997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04113147035241127, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.039555102586746216, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02738238498568535, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.023578833788633347, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.022244254127144814, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02192649058997631, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013857055455446243, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.011626571416854858, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011485274881124496, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010644383728504181, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.010502342134714127, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007393343839794397, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007349406369030476, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006997442338615656, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005049038212746382, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09883175790309906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08599887043237686, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07988628000020981, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07648708671331406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.068972647190094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03968258574604988, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03678740933537483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04720522090792656, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04336145520210266, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04028612747788429, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03532702103257179, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03388693183660507, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.023799749091267586, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02062140591442585, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.019007738679647446, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01861938089132309, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011897582560777664, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009786292910575867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009598089382052422, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008913391269743443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008711553178727627, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006241189781576395, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006128459703177214, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0057229334488511086, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003959163092076778, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08599887043237686, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2422860860824585, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22605422139167786, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22084936499595642, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19947907328605652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11309532821178436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10700073093175888, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1261952966451645, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11654723435640335, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.114711232483387, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10062172263860703, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09577275067567825, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06388911604881287, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.055540069937705994, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05398707091808319, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05360466614365578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.031856317073106766, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027302533388137817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027177099138498306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02471446618437767, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02462732046842575, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01630914956331253, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015814440324902534, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015741048380732536, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00949688907712698, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1261952966451645, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17363020777702332, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13455551862716675, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12100251764059067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0978236198425293, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07679896801710129, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06058060750365257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09246646612882614, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08475586026906967, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08052050322294235, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05262969434261322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05012698099017143, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04690774530172348, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.040363654494285583, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03680608421564102, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03593552112579346, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.023450125008821487, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01925048604607582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018983185291290283, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014635825529694557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014097483828663826, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012441889382898808, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01250295527279377, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011302467435598373, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008560248650610447, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17363020777702332, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21242810785770416, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1993832290172577, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19544556736946106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17767781019210815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09995176643133163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09517648816108704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1106572225689888, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10239096730947495, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10122010111808777, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08973971754312515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08569657057523727, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.056396231055259705, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04900375008583069, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.047912538051605225, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04765726998448372, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02820131927728653, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02474217116832733, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02467063069343567, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022682994604110718, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022639548406004906, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014917552471160889, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015050666406750679, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01457564253360033, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010173595510423183, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17767781019210815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2541947364807129, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23876450955867767, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23408973217010498, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21277298033237457, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11957624554634094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11388963460922241, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13220708072185516, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12243779748678207, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12108165770769119, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10738399624824524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10249845683574677, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0672803521156311, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05849143862724304, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.057195112109184265, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05688728392124176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03358849138021469, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02922207862138748, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029135406017303467, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026720693334937096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026671526953577995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017520317807793617, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017299050465226173, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01710011251270771, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011085379868745804, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13220708072185516, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.220644012093544, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2047363966703415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19680377840995789, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17926813662052155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09977929294109344, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09371251612901688, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11663052439689636, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10773953050374985, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10189768671989441, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09074005484580994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0869118794798851, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0586688369512558, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05109993740916252, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04761720448732376, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04678058251738548, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029318058863282204, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024605078622698784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024239080026745796, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022842207923531532, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02234172821044922, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015411523170769215, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01541328988969326, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014280742034316063, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010187337175011635, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17926813662052155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08577150106430054, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07959050685167313, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07640755921602249, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0690113753080368, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0399477481842041, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.037017107009887695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04696988686919212, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.043181441724300385, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04058673605322838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0354713499546051, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03398274630308151, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02382540889084339, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02065178193151951, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01922452077269554, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018873903900384903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011956039816141129, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010023503564298153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009863446466624737, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009133271872997284, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008962580002844334, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0063611953519284725, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00640825042501092, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005924853030592203, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004394086543470621, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08577150106430054, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07544399797916412, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06993857771158218, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06661097705364227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05998927727341652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.034780438989400864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03201158717274666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04170714691281319, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.038503531366586685, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03536935895681381, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03090381994843483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02956310845911503, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.021016031503677368, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.018322166055440903, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016693737357854843, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01629621349275112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010532508604228497, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00863033439964056, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008425941690802574, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007850377820432186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0076311733573675156, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005558463744819164, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00548703083768487, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005047420505434275, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003587398212403059, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07544399797916412, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2400275468826294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2237149477005005, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21851849555969238, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1973334550857544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1120859906077385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1059727668762207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12532667815685272, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1157194972038269, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11369512975215912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09971795976161957, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0948566198348999, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0634489357471466, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0551622211933136, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05348611995577812, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05310612544417381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03164622560143471, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027076421305537224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026946812868118286, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024519221857190132, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024414075538516045, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016210880130529404, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015746327117085457, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01561956387013197, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009484376758337021, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12532667815685272, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14443041384220123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1145818829536438, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10431554913520813, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08313052356243134, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06317532062530518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05109329894185066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07678767293691635, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07002463936805725, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06661663204431534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04576801136136055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04277603328227997, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03914331644773483, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03354373574256897, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.030501287430524826, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029750697314739227, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019686205312609673, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016347132623195648, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016129346564412117, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013152801431715488, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012774123810231686, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010828294791281223, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011136490851640701, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009866216219961643, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00822021160274744, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14443041384220123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21253646910190582, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19953982532024384, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19563482701778412, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17771635949611664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10002817958593369, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09523432701826096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11068449914455414, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10245693475008011, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10130096226930618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08978728204965591, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08571874350309372, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.056380853056907654, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04901896044611931, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04792807623744011, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04766865074634552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.028179500252008438, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02470000647008419, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0246269553899765, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022633718326687813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02259216271340847, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014845901168882847, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014956850558519363, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01448890846222639, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010024811141192913, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17771635949611664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25296685099601746, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23760968446731567, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2329799234867096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2117306888103485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1190168485045433, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11335086077451706, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13161729276180267, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1218489408493042, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12051080167293549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10687686502933502, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1020454540848732, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06700219213962555, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05822117254137993, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05694088712334633, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05663667991757393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03345656767487526, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029109295457601547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02902519330382347, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026621662080287933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026575272902846336, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017488770186901093, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01725667342543602, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017075685784220695, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011099481955170631, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13161729276180267, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21962454915046692, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20314395427703857, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19476136565208435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17731709778308868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0990855023264885, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09277762472629547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11682401597499847, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10768644511699677, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10141634196043015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09000188857316971, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08625777810811996, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.058787647634744644, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0510324165225029, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04731994494795799, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04641035944223404, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029364220798015594, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024461844936013222, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02407868206501007, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022690605372190475, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022159351035952568, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015520736575126648, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015392321161925793, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.0143305454403162, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010190226137638092, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17731709778308868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08852855116128922, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08227023482322693, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07904379814863205, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07131819427013397, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04114346206188202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.038261182606220245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.047944437712430954, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.044470224529504776, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04179749637842178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03660278022289276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.034933459013700485, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02425963245332241, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.021243544295430183, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019779914990067482, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.019422519952058792, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012150270864367485, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010311746038496494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010147877968847752, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00941004604101181, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009236995130777359, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006455229129642248, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006587801035493612, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006013540085405111, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004512852989137173, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08852855116128922, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08265449106693268, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07683773338794708, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07332967966794968, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06600931286811829, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.038182977586984634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03522336483001709, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04556336626410484, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04204531013965607, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03881507366895676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03396402671933174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.032513197511434555, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.022977199405431747, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02005842514336109, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.018325308337807655, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.017892690375447273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011514381505548954, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009462057612836361, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009256770834326744, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008609293028712273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008385458029806614, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00608640443533659, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005995768588036299, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005547795444726944, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003908785060048103, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08265449106693268, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2548159062862396, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23788973689079285, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23251666128635406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20997953414916992, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11926387250423431, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11284112930297852, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13315756618976593, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12287716567516327, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12094540894031525, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10618310421705246, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10104943811893463, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06747488677501678, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.058601707220077515, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05695940926671028, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.056570302695035934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.033662110567092896, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028836771845817566, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02870752103626728, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.026107193902134895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02602154202759266, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.017239954322576523, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016739455983042717, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016636541113257408, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.01010102964937687, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13315756618976593, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13707204163074493, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1146240308880806, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10391994565725327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08551404625177383, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.060907941311597824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.051552675664424896, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07727160304784775, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07081659138202667, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06387525051832199, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04735316336154938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04429531842470169, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0393516905605793, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03400329872965813, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02961389534175396, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028487680479884148, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019822848960757256, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01608183979988098, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01564832404255867, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013614419847726822, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01299339160323143, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010965807363390923, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011425377801060677, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009653417393565178, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008527937345206738, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13707204163074493, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2088284194469452, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19602645933628082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19215735793113708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17458616197109222, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09826770424842834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09355467557907104, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10862529277801514, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10065879672765732, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09951502084732056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08819643408060074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08415240794420242, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05524413287639618, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.048142820596694946, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04706680774688721, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04681273549795151, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027599534019827843, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02421371266245842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02414015121757984, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022174514830112457, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02213437482714653, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014434313401579857, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014595981687307358, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014082285575568676, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009701416827738285, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19215735793113708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24658571183681488, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2315724790096283, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22704249620437622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20637907087802887, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11600954830646515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1104561910033226, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12818370759487152, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11877774447202682, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11745818704366684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10414350777864456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09935243427753448, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06521885097026825, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05672641843557358, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.055467404425144196, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.055168990045785904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.032543521374464035, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028301605954766273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02822175808250904, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025865085422992706, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025818131864070892, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016913631930947304, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01668977364897728, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01650143601000309, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010605260729789734, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12818370759487152, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21775612235069275, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20057924091815948, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1918962448835373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17468880116939545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09806730598211288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0914706215262413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11588546633720398, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10692822188138962, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1005575954914093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08880596607923508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08513601869344711, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0583747923374176, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0506666898727417, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04681648686528206, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.045883093029260635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.0291360542178154, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024184929206967354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02379607781767845, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02237257920205593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021824030205607414, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01534191332757473, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01522174384444952, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01409999094903469, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010031069628894329, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1918962448835373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09855357557535172, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09177866578102112, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08856802433729172, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08002784103155136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0459417924284935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.042907655239105225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05294453725218773, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.049112360924482346, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04663762077689171, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.040919508785009384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.039065711200237274, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.026819352060556412, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.023462748154997826, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.022058164700865746, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.021713407710194588, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013432453386485577, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.011436455883085728, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011285670101642609, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010433388873934746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.010278078727424145, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007102083880454302, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007183301728218794, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006679264362901449, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0048411451280117035, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09855357557535172, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08845619112253189, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08217316120862961, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07868549227714539, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07083095610141754, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.040764909237623215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.037779513746500015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04829336330294609, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.044613711535930634, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.041412923485040665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.036280419677495956, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.034722261130809784, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02429208904504776, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.021226385608315468, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.019553879275918007, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.019130798056721687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01215791329741478, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.010055351071059704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009853130206465721, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.009152760729193687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008941104635596275, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006387207191437483, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006298172753304243, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005872522480785847, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004061045125126839, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08845619112253189, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2540622651576996, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23731759190559387, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2322167009115219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2098034918308258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11880604177713394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11254598200321198, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13241301476955414, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12219668179750443, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12039103358983994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10589145123958588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.1008811891078949, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0671297237277031, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05826989561319351, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.056692954152822495, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.056327901780605316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.033475324511528015, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028658783063292503, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.028538255020976067, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02599632553756237, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025903688743710518, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01712905988097191, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01656152866780758, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016543429344892502, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009912621229887009, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13241301476955414, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16717518866062164, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13707992434501648, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1245904415845871, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1062384843826294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07346870005130768, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0608782097697258, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09246747195720673, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08494194597005844, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07812679558992386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05795792117714882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.054687872529029846, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.047232527285814285, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.040784627199172974, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03561554104089737, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.034304019063711166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.023693392053246498, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01915127784013748, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01871119998395443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016270026564598083, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015532564371824265, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012932672165334225, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013375910930335522, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011323314160108566, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009764251299202442, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16717518866062164, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20788410305976868, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19506756961345673, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1912328600883484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17372173070907593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09781898558139801, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09312104433774948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10818102210760117, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10022138059139252, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.0990658551454544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08779681473970413, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08380010724067688, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.055069103837013245, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04793673753738403, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.046864524483680725, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04660845547914505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02752375788986683, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02414357103407383, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024074263870716095, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02212088741362095, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022077210247516632, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014467221684753895, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014606202021241188, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014116751961410046, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009775816462934017, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1912328600883484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2462720274925232, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23124052584171295, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22669707238674164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20601460337638855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11581209301948547, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11029423028230667, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1279354989528656, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11857713758945465, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11727644503116608, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10395713150501251, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0991751030087471, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06508893519639969, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.056629110127687454, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05537877604365349, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05508756265044212, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.032476820051670074, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028269615024328232, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028185464441776276, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025837339460849762, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025791490450501442, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016863562166690826, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016690798103809357, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016452902927994728, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010641876608133316, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1279354989528656, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21537253260612488, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19739852845668793, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18845069408416748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17122167348861694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09679869562387466, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08988730609416962, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11488054692745209, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10593607276678085, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09940651059150696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08725173771381378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08358583599328995, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0578451082110405, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.050152990967035294, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04621139168739319, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.045249056071043015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028869371861219406, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0238775834441185, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023489344865083694, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022014526650309563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021451734006404877, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015176964923739433, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015062255784869194, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013883674517273903, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009932358749210835, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18845069408416748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09690637141466141, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09031368046998978, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08737565577030182, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.078915074467659, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04519389569759369, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04233165457844734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.051988378167152405, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04796907305717468, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04585809260606766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04025310277938843, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03847074881196022, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.026324128732085228, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02292102575302124, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.021688202396035194, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.021393226459622383, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013195609673857689, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.011209413409233093, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011081180535256863, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010218988172709942, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.010088350623846054, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006964745931327343, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006958856713026762, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0065877134911715984, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004637380596250296, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09690637141466141, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08651228249073029, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08036041259765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07726988196372986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06961306184530258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03998531401157379, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03718329966068268, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04702747240662575, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04325549304485321, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04060631990432739, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03559635207056999, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03410445898771286, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.023702440783381462, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.020597442984580994, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01916848123073578, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.018807727843523026, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011853963136672974, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009851641952991486, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009682504460215569, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008963776752352715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00878515187650919, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006200645584613085, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006111960392445326, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0057455324567854404, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0039347014389932156, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08651228249073029, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24193882942199707, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22593146562576294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22095273435115814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1996874064207077, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1130714938044548, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10708373039960861, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12593765556812286, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11636506766080856, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11462169140577316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10074837505817413, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09585796296596527, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06381220370531082, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.055434614419937134, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.053964704275131226, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05361807718873024, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03181293234229088, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02728370949625969, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027169235050678253, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02471921592950821, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024642622098326683, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016277719289064407, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015756236389279366, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01573246717453003, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009423806332051754, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12593765556812286, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16518323123455048, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13212677836418152, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12207119166851044, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10423668473958969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07351508736610413, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06159508228302002, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08612431585788727, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07873222231864929, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07658569514751434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05408448353409767, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05157989263534546, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.044002730399370193, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03807450458407402, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03575742244720459, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03519978001713753, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022358817979693413, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01949695684015751, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019371958449482918, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016000129282474518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01571490615606308, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012716026976704597, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013479613699018955, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012042323127388954, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010498850606381893, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16518323123455048, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2034522294998169, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1909320056438446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1871652603149414, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17001742124557495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09572732448577881, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09113053232431412, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10585195571184158, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09805939346551895, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09695310890674591, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0859217420220375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08196518570184708, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.053859513252973557, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.046897921711206436, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.045848920941352844, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.045596856623888016, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.026946168392896652, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02359868586063385, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02353091351687908, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02161400206387043, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.021571213379502296, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014160778373479843, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014238644391298294, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013817215338349342, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009479493834078312, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1909320056438446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23800992965698242, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2234869748353958, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2191367745399475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19912339746952057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11195351928472519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10658741742372513, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12366469204425812, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11464237421751022, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11337368190288544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10051481425762177, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09583855420351028, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0628826841711998, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.054745666682720184, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.053527649492025375, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05324087291955948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.031383220106363297, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02731793001294136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027238233014941216, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024965256452560425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024921366944909096, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01628170721232891, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016116607934236526, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015882929787039757, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010256837122142315, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12366469204425812, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20358683168888092, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18596580624580383, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.17704540491104126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16117514669895172, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09147658199071884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0845482274889946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10938609391450882, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10071348398923874, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09415477514266968, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08238386362791061, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07900159806013107, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05496154725551605, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.047805726528167725, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04378512501716614, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04278439283370972, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027534035965800285, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022770380601286888, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022361112758517265, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02100929245352745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02042027935385704, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014672897756099701, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014613124541938305, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.013399748131632805, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009847981855273247, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18596580624580383, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09185287356376648, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08541753143072128, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08236300200223923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0742354691028595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04268965497612953, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03984709829092026, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04974238947033882, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04572918266057968, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.043344538658857346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0379825122654438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03631534054875374, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.025246039032936096, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.021898655220866203, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.020542830228805542, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02020811103284359, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012671137228608131, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010733097791671753, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010582436807453632, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009800524450838566, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009653891436755657, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006752851419150829, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006858015898615122, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0063094994984567165, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004756255075335503, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09185287356376648, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08270280063152313, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07686107605695724, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07370886206626892, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06635502725839615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.038112033158540726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03537491708993912, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04481983929872513, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04141485318541527, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.038692906498909, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0339067205786705, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03236974775791168, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.022594181820750237, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019713999703526497, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.018258266150951385, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.017903557047247887, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011312643066048622, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009408887475728989, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00924072414636612, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008561895228922367, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008391511626541615, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005950885824859142, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005883440840989351, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005504094064235687, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0038319495506584644, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08270280063152313, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22275055944919586, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20791849493980408, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20321807265281677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18354853987693787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10390321165323257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09834303706884384, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11608295142650604, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10714967548847198, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10534091293811798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09257683902978897, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08814360201358795, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05881005525588989, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05107016861438751, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04959900304675102, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04924050718545914, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02933564782142639, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025104945525527, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024988949298858643, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022763703018426895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02266078069806099, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01502482034265995, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014588363468647003, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014482189901173115, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008807164616882801, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18354853987693787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14245764911174774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11311779916286469, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1032259538769722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08688373863697052, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06261610239744186, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.051568303257226944, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07624194025993347, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06946056336164474, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06623122096061707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0463491752743721, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04403151944279671, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03898089751601219, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03362341597676277, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.030605299398303032, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029862288385629654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019739747047424316, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016912611201405525, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01672649011015892, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014047902077436447, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013643410056829453, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.011196707375347614, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012085383757948875, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010295126587152481, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009522726759314537, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14245764911174774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19901324808597565, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18677201867103577, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18307824432849884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.16641338169574738, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09368078410625458, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08920757472515106, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10371670871973038, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09596886485815048, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09487367421388626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08411611616611481, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08033454418182373, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05292361602187157, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.045977603644132614, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04495261237025261, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04470960423350334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.026449020951986313, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.023278046399354935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02321034111082554, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.021358292549848557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.021319253370165825, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014030810445547104, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014257876202464104, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013704192824661732, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009765181690454483, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18677201867103577, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23306980729103088, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.21883447468280792, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.21456684172153473, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19499601423740387, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10966792702674866, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10443665832281113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.121148981153965, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1122860535979271, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11105064302682877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09846662729978561, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09392410516738892, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06164705008268356, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05366184934973717, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.052475783973932266, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.052196234464645386, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03078172355890274, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.026893481612205505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.026814507320523262, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02460993453860283, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024566566571593285, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016056818887591362, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016044367104768753, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015671683475375175, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01045215968042612, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.121148981153965, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20369863510131836, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18548937141895294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.17644518613815308, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16046379506587982, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09147310256958008, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08432336151599884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10961306840181351, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10073620080947876, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09416645020246506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08208485692739487, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0787442997097969, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05517951771616936, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.047813475131988525, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04378339275717735, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.042793434113264084, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027587244287133217, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022809388116002083, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02240964025259018, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.020995017141103745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.020413659512996674, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014629200100898743, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014683172106742859, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01332138106226921, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.00996953435242176, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18548937141895294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08977837860584259, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0835828185081482, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08069997280836105, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07286320626735687, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04177143797278404, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03905945643782616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04827886447310448, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04455634951591492, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.042401496320962906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03721665218472481, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.035466406494379044, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0244577806442976, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.021307693794369698, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.020063133910298347, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0197650957852602, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012265753000974655, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01042655948549509, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010293742641806602, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009518666192889214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009384753182530403, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006510511040687561, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006570043973624706, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006136616226285696, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004473715554922819, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08977837860584259, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08430764824151993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07838792353868484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07516604661941528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06771783530712128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03884634003043175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03606404736638069, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04564899578690529, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.042188044637441635, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03943520039319992, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03457295522093773, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03302765637636185, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.023002073168754578, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02009832113981247, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01863167993724346, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.018257683143019676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01152640301734209, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009603243321180344, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009428467601537704, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00874802190810442, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008562054485082626, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006071086041629314, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006012288853526115, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0056171538308262825, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003925871569663286, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08430764824151993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22926971316337585, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21405772864818573, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2092968225479126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18901300430297852, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10688454657793045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10129284113645554, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11905063688755035, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11000030487775803, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1083611398935318, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09527557343244553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09069318324327469, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06032786890864372, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0524090938270092, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05099666863679886, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.050682783126831055, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030085165053606033, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025892792269587517, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025796230882406235, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023515651002526283, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023429663851857185, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015458853915333748, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01514514721930027, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014949755743145943, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009336269460618496, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18901300430297852, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14767180383205414, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11487141996622086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1011333167552948, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08321245014667511, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06229414418339729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04944518208503723, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08297496289014816, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07545310258865356, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06865882128477097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04711960256099701, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04530945420265198, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.042476579546928406, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03636143356561661, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03046049177646637, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028905820101499557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.021552113816142082, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01680925115942955, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01636691763997078, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014123938977718353, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013209912925958633, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012116854079067707, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012424984946846962, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010338188149034977, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009408265352249146, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14767180383205414, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18596890568733215, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17453306913375854, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1710592657327652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.15545925498008728, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08746123313903809, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08326402306556702, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09669312834739685, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08960743248462677, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08858010172843933, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07851683348417282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07496462762355804, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04922250285744667, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.042867790907621384, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04190347343683243, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.041673533618450165, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.024591656401753426, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.021599700674414635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.021537411957979202, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019796183332800865, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.019757162779569626, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012897948734462261, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01308539230376482, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012581362389028072, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008776366710662842, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18596890568733215, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.22163240611553192, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2081068754196167, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.20405249297618866, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18547660112380981, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10425788164138794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09929543733596802, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11526346206665039, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10677331686019897, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10558649897575378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09361840039491653, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08936657011508942, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05866404250264168, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.051031384617090225, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04989834129810333, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04963144287467003, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.029316790401935577, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02559804916381836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025522656738758087, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02343130111694336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023386280983686447, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015346714295446873, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015310540795326233, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014968767762184143, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010022739879786968, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18547660112380981, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.18088264763355255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.16304318606853485, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.15395322442054749, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.13997744023799896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.08131667226552963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.07395217567682266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09905911237001419, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09068964421749115, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08403228223323822, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0724196806550026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.06956425309181213, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05005316808819771, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04363693669438362, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.039475057274103165, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.03843075782060623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.025277724489569664, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.021474791690707207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02103331685066223, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.019854864105582237, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.019233442842960358, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01389237493276596, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014999845996499062, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.012463079765439034, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011364217847585678, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.18088264763355255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09785553812980652, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09120204299688339, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08827310800552368, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07968928664922714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04546694830060005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04261426255106926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05241239443421364, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.048214465379714966, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.046116236597299576, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0405682772397995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03874877467751503, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02655881829559803, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02303016558289528, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.021794456988573074, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.021498490124940872, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013288285583257675, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01126572210341692, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011136340908706188, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010284884832799435, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.010154301300644875, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006995540577918291, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00699628284201026, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006598045118153095, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004670406226068735, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09785553812980652, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09001001715660095, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08382268249988556, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0805032029747963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07255679368972778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0414123460650444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03854075446724892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.048460718244314194, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04491833969950676, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04201769083738327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.036934878677129745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03521363437175751, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.024373957887291908, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02134896069765091, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.019813131541013718, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01943359710276127, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.012186919339001179, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.010152477771043777, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009966697543859482, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.009242335334420204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.009053836576640606, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006367004010826349, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006267451215535402, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0058906120248138905, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003972487524151802, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09001001715660095, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23287439346313477, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21767815947532654, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2132764309644699, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19287267327308655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10844933241605759, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10301431268453598, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12065020203590393, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11133288592100143, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10989546775817871, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09692440927028656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0922829657793045, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06105707585811615, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05304984003305435, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.051758669316768646, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05143756419420242, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0304875411093235, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026268862187862396, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026173945516347885, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023891430348157883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023828379809856415, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01566254161298275, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015317317098379135, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01517943199723959, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009427466429769993, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19287267327308655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1778244823217392, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13974040746688843, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.126247838139534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10262662172317505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07895060628652573, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06422564387321472, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0961175262928009, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08723414689302444, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08285833895206451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05677120015025139, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05178603529930115, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04916353151202202, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04199293255805969, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03828790783882141, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03737996518611908, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024716727435588837, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020678427070379257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020396430045366287, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016621045768260956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01613095961511135, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013594942167401314, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014242126606404781, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012457594275474548, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01069161668419838, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1778244823217392, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1694091409444809, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.15897385776042938, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.15583427250385284, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.14159414172172546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07972007244825363, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07591883838176727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08822257816791534, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08166356384754181, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08071426302194595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0716104507446289, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06844118982553482, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04505876451730728, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03927459940314293, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.038407355546951294, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03819695860147476, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.022538620978593826, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.020186765119433403, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.020129498094320297, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01859882101416588, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.018565624952316284, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01208529993891716, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012805670499801636, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01180477999150753, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009282354265451431, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1694091409444809, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.19951900839805603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18740594387054443, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.18372265994548798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1670714169740677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09413307160139084, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0896870344877243, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10410667210817337, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09640505164861679, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09532571583986282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08462758362293243, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08090592920780182, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.053284429013729095, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04649066552519798, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04548458755016327, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04524197429418564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.026667991653084755, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024147605523467064, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024080786854028702, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02231503278017044, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022276511415839195, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014443183317780495, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01562648080289364, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014119715429842472, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011667432263493538, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.18740594387054443, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.163803830742836, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.14718976616859436, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1357654631137848, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.12318471819162369, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.07345634698867798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0653928816318512, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09573882073163986, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.08659256249666214, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.07591157406568527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.06513581424951553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.06279955059289932, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04737354442477226, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.041195377707481384, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.03586381673812866, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.034531500190496445, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.023899193853139877, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.019863246008753777, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.019308917224407196, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.0183878131210804, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.01765913888812065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013352985493838787, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01445749867707491, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.011799218133091927, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01124753337353468, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.163803830742836, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08803928643465042, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08201892673969269, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07928186655044556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0716947391629219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.040895283222198486, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03830708935856819, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.047297585755586624, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04349830746650696, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04148173704743385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03647707402706146, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03501887619495392, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0239774938672781, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.020785417407751083, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019636712968349457, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01935887336730957, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012031102553009987, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010209922678768635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01008480042219162, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009335339069366455, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009212649427354336, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006402974482625723, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006435930263251066, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006060177925974131, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004403506405651569, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08803928643465042, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0801510214805603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07450845837593079, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07154213637113571, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06450734287500381, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03685186430811882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0342552587389946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04317012429237366, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.039949897676706314, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.037397243082523346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03280244767665863, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.031339600682258606, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.021776696667075157, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01903449557721615, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.017651136964559555, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01731528341770172, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010900411754846573, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009113305248320103, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008952977135777473, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00830664113163948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008133966475725174, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005757533945143223, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005713308695703745, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005336315371096134, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0037449293304234743, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0801510214805603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2215978503227234, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20709528028964996, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20266221463680267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18327756226062775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10305356979370117, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09781784564256668, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11476846784353256, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10595589131116867, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10447150468826294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09206537157297134, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0877140611410141, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05813261866569519, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05047299340367317, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.049179837107658386, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.048855531960725784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02897617779672146, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02490568719804287, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024807972833514214, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02262815274298191, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02255542203783989, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014871758408844471, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01446530967950821, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014398366212844849, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008793767541646957, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18327756226062775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.15170258283615112, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1022571325302124, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08129668235778809, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07789008319377899, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06579755991697311, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04149560257792473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08590717613697052, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07786428183317184, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07091861963272095, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04401690140366554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04423997551202774, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04397037252783775, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03758920729160309, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03227592632174492, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.030897630378603935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02233726903796196, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017938755452632904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017500940710306168, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014185434207320213, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013301028870046139, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012720937840640545, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013222873210906982, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011164268478751183, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010236520320177078, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.15170258283615112, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15493781864643097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.14546966552734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.14258240163326263, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.129636749625206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07268411666154861, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06921327114105225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08054228127002716, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07449837028980255, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.0735953226685524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0653216615319252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0624069981276989, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04098598286509514, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.035662904381752014, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03484579548239708, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03465171158313751, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02049136534333229, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.018038513138890266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.017980922013521194, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01656474359333515, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.016530808061361313, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.010823492892086506, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.011047412641346455, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0105522396042943, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.007558026816695929, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15493781864643097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17725485563278198, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1665511280298233, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.16327325999736786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.14853747189044952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08335188776254654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07941513508558273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09241756796836853, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.0853847861289978, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08438658714294434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07494550943374634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0716932862997055, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04721616208553314, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.040989745408296585, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04007565602660179, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03986012935638428, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.023623455315828323, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02095191366970539, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.020892761647701263, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.019297413527965546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.019262779504060745, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012693065218627453, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01312621496617794, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012399444356560707, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009335298091173172, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17725485563278198, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.13381725549697876, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.11869899183511734, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1090940311551094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.09841538965702057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.06047184765338898, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.05312637239694595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.0806020051240921, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.07122548669576645, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.06266000121831894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.053159695118665695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.05146627873182297, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.040489524602890015, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.03507791832089424, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.03035353496670723, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.029115043580532074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.021060632541775703, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.01798866130411625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.01746220700442791, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.016810961067676544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.016141656786203384, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01264062337577343, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014196876436471939, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.011232580989599228, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011930776759982109, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.13381725549697876, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08711551874876022, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08120466768741608, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0786338597536087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07110219448804855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04037974029779434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.037873003631830215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04627581313252449, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.042772114276885986, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.040964506566524506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03603525832295418, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03443138673901558, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.023422524333000183, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02042362280189991, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019362766295671463, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.019102463498711586, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011737333610653877, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010023588314652443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009913302026689053, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009160601533949375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009047525934875011, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0062147811986505985, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006251181475818157, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005895607639104128, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00421687588095665, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08711551874876022, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07879360765218735, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0733323022723198, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07060835510492325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06374819576740265, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03627390041947365, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03380647301673889, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.042237788438797, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03904466703534126, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.036806635558605194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03232870623469353, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.030920935794711113, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02132595330476761, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.018608249723911285, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01736508682370186, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.017063554376363754, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010669760406017303, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008918353356420994, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008774399757385254, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008127023465931416, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007978031411767006, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005598919466137886, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005518370307981968, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005224957130849361, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003550038207322359, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07879360765218735, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23767156898975372, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2225627601146698, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21787066757678986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19715335965156555, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11081273853778839, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10520168393850327, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12340401113033295, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11381655186414719, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11228738725185394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09911104291677475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0944070965051651, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06251224875450134, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05425535887479782, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05290825292468071, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05258210003376007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.031192945316433907, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026857011020183563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02676611766219139, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02443874068558216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024385785683989525, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016049223020672798, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015693970024585724, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015551356598734856, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009717991575598717, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12340401113033295, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14043545722961426, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11101196706295013, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09918564558029175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08223526179790497, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.058996137231588364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04685533419251442, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0798499584197998, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0709010660648346, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06405165791511536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04620595648884773, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.043806519359350204, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03993122652173042, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.034586988389492035, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.029343534260988235, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.027985097840428352, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020786285400390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017112258821725845, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016704540699720383, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015023486688733101, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01431978028267622, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01261057797819376, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013357582502067089, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011207295581698418, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010956116952002048, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14043545722961426, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.13324368000030518, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.12519057095050812, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.12269224971532822, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.11163118481636047, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.06266950070858002, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.05974220111966133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.06949939578771591, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.06422343105077744, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.06345166265964508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.05643071234226227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.05403968319296837, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.03568439558148384, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.031092951074242592, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03040810115635395, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03024778887629509, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.017920028418302536, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.016367169097065926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.016324564814567566, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.015187826007604599, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.015162848867475986, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.009922684170305729, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.010878641158342361, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.009714650921523571, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.00840306282043457, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.13324368000030518, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14994068443775177, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1408928483724594, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.13810290396213531, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1257365494966507, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0703212171792984, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06700775772333145, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.07818183302879333, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07208368927240372, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07119385153055191, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06328777223825455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06064425781369209, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.03997501730918884, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03458964079618454, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03379972279071808, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0336153581738472, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02012685313820839, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.017675261944532394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.01762421801686287, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.016291305422782898, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01625795289874077, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.010970051400363445, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.011073693633079529, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.010704269632697105, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.007875390350818634, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14994068443775177, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.09514833241701126, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.08402489125728607, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.07584339380264282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.0683833435177803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.043205372989177704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.037087827920913696, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.061560504138469696, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.052707698196172714, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.04500773921608925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.037894200533628464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.03731444850564003, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.031221676617860794, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.026055486872792244, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.021819451823830605, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.02068479359149933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.01618298329412937, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.013043126091361046, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.012561517767608166, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.01219497062265873, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.011586717329919338, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.009521862491965294, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01049700565636158, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.008034844882786274, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.008802005089819431, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.09514833241701126, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07731602340936661, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07191678136587143, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06947275251150131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0627039298415184, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03577784448862076, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03345305100083351, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.041434697806835175, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03813673555850983, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03631152957677841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.031846173107624054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030528748407959938, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020975811406970024, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018234405666589737, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017197715118527412, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01694450154900551, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010533155873417854, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008991275914013386, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008881695568561554, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00821988470852375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008107217960059643, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0056444681249558926, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005741065833717585, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005337224341928959, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0040094261057674885, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07731602340936661, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06813088059425354, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06320268660783768, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06092014163732529, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05486325919628143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03135897219181061, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0292039904743433, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03636907786130905, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03362926468253136, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03183041140437126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02784009277820587, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026548685505986214, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018350550904870033, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016026638448238373, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015024974942207336, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01478369627147913, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009201863780617714, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007777602877467871, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007665259297937155, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007079113740473986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006959410384297371, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00488428445532918, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004877234809100628, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004583798814564943, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032644763123244047, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06813088059425354, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20874451100826263, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1949569284915924, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1904207170009613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17217116057872772, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09727419167757034, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09208498895168304, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10883799940347672, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10035032033920288, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09864545613527298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08675120025873184, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08265101164579391, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05516541004180908, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04790656268596649, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.046528272330760956, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04619254544377327, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02753954567015171, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023734666407108307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023616639897227287, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02155931480228901, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021482540294528008, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014234579168260098, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014077555388212204, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013746719807386398, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008943183347582817, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1904207170009613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09649109840393066, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06449326872825623, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05264141038060188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04694870114326477, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04044598713517189, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0260199923068285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05243569612503052, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.047319695353507996, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0448031947016716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.027334949001669884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.026555776596069336, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02696358971297741, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02333056926727295, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02030888758599758, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01954628899693489, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013898800127208233, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.012035660445690155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011898459866642952, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010027149692177773, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009604389779269695, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008310230448842049, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009504682384431362, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007478215731680393, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008049814961850643, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09649109840393066, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.10628265887498856, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.0995098352432251, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.09736524522304535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.08837999403476715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.04955088347196579, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.047051120549440384, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.05516479164361954, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.050962984561920166, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.05020947381854057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.04438839852809906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0424138605594635, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.028091618791222572, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0244064349681139, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0237788874655962, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.023630278185009956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.01404986809939146, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.012392043136060238, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.01234673336148262, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.011373011395335197, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.011338512413203716, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.007478180807083845, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.007727312855422497, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.007273541763424873, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.005424868781119585, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.10628265887498856, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.10923886299133301, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.10215319693088531, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.09991724044084549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.09063215553760529, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.05101001635193825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.04840756207704544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.05702206864953041, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.05253256484866142, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.05170595273375511, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.04568345472216606, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.04371972754597664, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.029195159673690796, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.025278273969888687, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.02460857294499874, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.024453943595290184, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.014672065153717995, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.013056403957307339, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.013007727451622486, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.012034534476697445, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.011994224973022938, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.008061423897743225, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.008456705138087273, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.007847827859222889, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.006284525617957115, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.10923886299133301, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.02535756677389145, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.021579505875706673, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.019004808738827705, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.01705298386514187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.011536664329469204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.009535809978842735, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.01610177755355835, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.014505917206406593, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.012185996398329735, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.010038625448942184, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.009774655103683472, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.00849139504134655, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.00755373714491725, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.006230752449482679, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.005875581409782171, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.004649253096431494, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.00421887869015336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.004096231423318386, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.004020008258521557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.00386092159897089, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.0030787107534706593, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.003730014432221651, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.002744116587564349, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.0033878348767757416, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.02535756677389145, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + } + ], + "last_module_idx": 162, + "base_perplexity": 5.687353312138993 +} \ No newline at end of file