diff --git "a/measurement.json" "b/measurement.json" new file mode 100644--- /dev/null +++ "b/measurement.json" @@ -0,0 +1,88612 @@ +{ + "measurement": [ + { + "key": "model.layers.0.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.0164031982421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0162353515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.007686614990234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0076446533203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0234832763671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.01611328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.007656097412109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.007625579833984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.008026123046875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.00909423828125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0076141357421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.00567626953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0048065185546875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0048065185546875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.004863739013671875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.004596710205078125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0046234130859375, + "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": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.007686614990234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.0171051025390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.016876220703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.008056640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.00799560546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0207366943359375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.016693115234375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0080413818359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.00795745849609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.008209228515625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.00881195068359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0079345703125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.004871368408203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0037994384765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.003795623779296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0035686492919921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0030517578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00339508056640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0037994384765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.055938720703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0465087890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.031524658203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.02496337890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.04534912109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.035430908203125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0265655517578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0204925537109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0211181640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.024261474609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.017364501953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01216888427734375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.00699615478515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.006229400634765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.006336212158203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.003742218017578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00420379638671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.003742218017578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.0243682861328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.01409912109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.009857177734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.00955963134765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0151824951171875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.01363372802734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.011474609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0068359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.00728607177734375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.00774383544921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0067291259765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.004047393798828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.00328826904296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.00292205810546875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.002471923828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00201416015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0025272369384765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00201416015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.05804443359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.054473876953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.048614501953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.025848388671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.02874755859375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.0266265869140625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.0261688232421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.023681640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0226287841796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.01459503173828125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0128173828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.007354736328125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.006847381591796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.006465911865234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.004138946533203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.00403594970703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.003772735595703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.004138946533203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.08868408203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.08343505859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.07452392578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.03955078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.04388427734375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.04052734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.03997802734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0361328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.034515380859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0221099853515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.01922607421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0110321044921875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.00977325439453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0091400146484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.00589752197265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0057373046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00428009033203125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.00589752197265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.057861328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.0491943359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.02923583984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.0246124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.04937744140625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.042999267578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.0260772705078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0218658447265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.021820068359375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.0220794677734375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.0200958251953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.01203155517578125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.00931549072265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0089874267578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.008697509765625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.00707244873046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.007129669189453125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.00707244873046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.01336669921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.01065826416015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.007183074951171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.005725860595703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.01044464111328125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.00832366943359375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.006214141845703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.004726409912109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.005046844482421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.005523681640625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.004100799560546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0028476715087890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0016374588012695312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0014715194702148438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0015325546264648438, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0009641647338867188, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0010023117065429688, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0009641647338867188, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.01258087158203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.00981903076171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.006656646728515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.00525665283203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0085906982421875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.007537841796875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0057830810546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.004238128662109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.004283905029296875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.004360198974609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0036163330078125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.002201080322265625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0014715194702148438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0012979507446289062, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0011949539184570312, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.000835418701171875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0008845329284667969, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.000835418701171875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.07763671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.06549072265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.052032470703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.03466796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.042999267578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.039398193359375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.036468505859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0280609130859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.026275634765625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.02197265625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.018951416015625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01102447509765625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0088653564453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.007579803466796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00582122802734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.005062103271484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004169464111328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004169464111328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.14404296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.086181640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.061492919921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06060791015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.08782958984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.07867431640625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.069091796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0396728515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.042510986328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.045440673828125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.03900146484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0233001708984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.018585205078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0153961181640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01332855224609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01087188720703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01271820068359375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0153961181640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.11126708984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.10595703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.0958251953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.049102783203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0540771484375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.049957275390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.04949951171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.045745849609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.04412841796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.027374267578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0238189697265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.01373291015625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0125579833984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.01197052001953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.00762939453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.007518768310546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0063629150390625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.00762939453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1326904296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.126708984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.11505126953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.05877685546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0645751953125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.0596923828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.05926513671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0548095703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.052764892578125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0323486328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.028106689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0160980224609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0140838623046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0133209228515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.00836181640625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.00823211669921875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00524139404296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.00836181640625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.11431884765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.10223388671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.0828857421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.053680419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.06915283203125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.0677490234375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.055084228515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0496826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.04913330078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.03826904296875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.038787841796875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.0164337158203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.014373779296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.01372528076171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.00927734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.007587432861328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0090484619140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.007587432861328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.0307464599609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.02838134765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.024505615234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.013824462890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0166473388671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.01488494873046875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0140838623046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0123291015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.01175689697265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.00848388671875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.00711822509765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.004299163818359375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.00350189208984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0032100677490234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.002330780029296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.002105712890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0016298294067382812, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0016298294067382812, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.0244903564453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0225677490234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0194244384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0109710693359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0128936767578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.01175689697265625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.011199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0097503662109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0092315673828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.00649261474609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.005588531494140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0032558441162109375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0027332305908203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0024967193603515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0017290115356445312, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0016012191772460938, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0011987686157226562, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0017290115356445312, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.11199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.10479736328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.09283447265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.051177978515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.057098388671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.052734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.051910400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.046173095703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.043853759765625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0288848876953125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.02508544921875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0143890380859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0124053955078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01137542724609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00742340087890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00716400146484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00457763671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00742340087890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1588134765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.126953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0888671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.070556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.08782958984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.07952880859375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.07525634765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.050933837890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.046600341796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.04522705078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0389404296875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0231475830078125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0198822021484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.016571044921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01345062255859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.012298583984375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01214599609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.016571044921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.15087890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.143798828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1307373046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.068359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.07537841796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.06951904296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.06903076171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0634765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.061431884765625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.038177490234375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0330810546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.01910400390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.01708984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.01617431640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.010406494140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01024627685546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.007770538330078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.007770538330078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.17626953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.16796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1527099609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0799560546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.088134765625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.081298828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08062744140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07427978515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.07177734375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.044525146484375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.03851318359375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.022186279296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.019378662109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.018280029296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01153564453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00742340087890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00742340087890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.167236328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.1519775390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.074462890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.086669921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.07958984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.07666015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.06646728515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.06378173828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.0438232421875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.037841796875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.02197265625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.01861572265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.017059326171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0118865966796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01116180419921875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.00830841064453125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.00830841064453125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.047149658203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.043853759765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.037811279296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0215301513671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0253448486328125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.022705078125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.021881103515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0189971923828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0182037353515625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0129852294921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.010894775390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0065765380859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.005428314208984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.004894256591796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0035228729248046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.003360748291015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0023708343505859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0035228729248046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.03741455078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.03472900390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0300445556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0169525146484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.01983642578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.01812744140625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.017242431640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.01514434814453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.01445770263671875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0100250244140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.008636474609375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.00502777099609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.00426483154296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.003910064697265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0026798248291015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00251007080078125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.001926422119140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.001926422119140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1573486328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1473388671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.130615234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0723876953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0811767578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0745849609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.07342529296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.06524658203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0621337890625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.04107666015625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.03546142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0205078125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0175323486328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01611328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01062774658203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01025390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0064697265625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0064697265625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.135498046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.11517333984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0841064453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.061248779296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.086669921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.07354736328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0638427734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.049072265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04534912109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.044158935546875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.036773681640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.023681640625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01922607421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.017333984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01477813720703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01302337646484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01351165771484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01477813720703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1710205078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1622314453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1463623046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.079345703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0875244140625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.0809326171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08013916015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0728759765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.070068359375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0447998046875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.038970947265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0224609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.020538330078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.01934814453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0124053955078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01220703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01038360595703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0124053955078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.205322265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1949462890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.176025390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.09515380859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.10504150390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.09698486328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.0960693359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0872802734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.08380126953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.05328369140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04620361328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0266265869140625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02349853515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0219573974609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01416778564453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.013885498046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00975799560546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01416778564453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.042022705078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.03900146484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.01751708984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.01123046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.0357666015625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.033111572265625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.0117034912109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.00949859619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.00894927978515625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.00710296630859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.006122589111328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.0036449432373046875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.00391387939453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0036640167236328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.00319671630859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0018978118896484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0017042160034179688, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0018978118896484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.06121826171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.056304931640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0487060546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.027862548828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0321044921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0295867919921875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0284271240234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0246734619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.023345947265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0162353515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0140838623046875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.00811004638671875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.00688934326171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.006256103515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.004245758056640625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0039825439453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0028781890869140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0039825439453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.053436279296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.049072265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.04241943359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0242919921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0282440185546875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.025909423828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0247650146484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.021514892578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0204620361328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0142974853515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0123443603515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.00714874267578125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.006008148193359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.00545501708984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.003734588623046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0034961700439453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00247955322265625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.003734588623046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1702880859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1588134765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.14013671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0787353515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.08831787109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.08148193359375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.080078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.070556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.06683349609375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.04473876953125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.038787841796875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.022308349609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01910400390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.017425537109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01148223876953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01105499267578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00701904296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00701904296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.181396484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.158935546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1239013671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.08447265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.10174560546875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.09295654296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.08734130859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0679931640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0611572265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.05279541015625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.045379638671875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0267791748046875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0227508544921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.019744873046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0150604248046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0138397216796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0125274658203125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0150604248046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.189697265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1795654296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1611328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.088134765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.09722900390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.0899658203125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08917236328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0804443359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0767822265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.04937744140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04296875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.024658203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0216827392578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0201416015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.012939453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0126495361328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.008758544921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.012939453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2216796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2098388671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.188720703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10308837890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11370849609375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.10528564453125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10430908203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.093994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.08978271484375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.057586669921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05010986328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0287322998046875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0249481201171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0231170654296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.014862060546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0145416259765625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.009063720703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.014862060546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2181396484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.197509765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.169921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.09930419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1143798828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.10552978515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.10198974609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.08697509765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08203125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.058135986328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.050323486328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.0290985107421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0247650146484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0223236083984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.015594482421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01471710205078125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.010772705078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.015594482421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.07257080078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.06671142578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.057525634765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.03314208984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.038330078125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.035400390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.033843994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0293121337890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0276947021484375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0194091796875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.01690673828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.00970458984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0082855224609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0075225830078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.005115509033203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.004787445068359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00362396240234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00362396240234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.062286376953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.057220458984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.04925537109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0284423828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0333251953125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.030609130859375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0290374755859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0251617431640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.023895263671875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0168609619140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0145721435546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0084381103515625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.007083892822265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0064239501953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.004425048828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.004116058349609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0030193328857421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.004116058349609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1815185546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1690673828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.148681640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.084228515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0946044921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.08721923828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.08551025390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0751953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.07110595703125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0479736328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.04156494140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0239410400390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02044677734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01861572265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.012359619140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01184844970703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00756072998046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00756072998046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.187744140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1650390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1314697265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.08746337890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.10186767578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.09307861328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.08990478515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0711669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.06378173828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.05267333984375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.045440673828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0267181396484375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02349853515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.020477294921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0151214599609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01436614990234375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0128631591796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0151214599609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.176025390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.165771484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.148681640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.08203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.09075927734375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08392333984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08306884765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07452392578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.071044921875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0462646484375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0401611328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0230865478515625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.020416259765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0189361572265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01227569580078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0119781494140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0087738037109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01227569580078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2275390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2147216796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1922607421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10614013671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11724853515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.10858154296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.107421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.09637451171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09185791015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.05950927734375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0517578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0296783447265625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0257568359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.02374267578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015380859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01500701904296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.009368896484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015380859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2210693359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.19970703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.171142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.100830078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1163330078125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.10723876953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.103759765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0880126953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0828857421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.05926513671875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.05126953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.0297698974609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0253753662109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0228424072265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0160980224609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01519775390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01136016845703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0160980224609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.08868408203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0823974609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0718994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.04095458984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.04669189453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.043121337890625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.041656494140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.036590576171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.034637451171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.023681640625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0206146240234375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01184844970703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.010223388671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0093536376953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.006256103515625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0059356689453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00441741943359375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.006256103515625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.07330322265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0682373046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.05963134765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.033905029296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.035858154296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.034454345703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0302886962890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.028839111328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0197906494140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.017120361328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.00991058349609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.00839996337890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.007671356201171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00519561767578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0048980712890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0034809112548828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0034809112548828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.198486328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1859130859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.164306640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0924072265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.10357666015625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0955810546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.09393310546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0830078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0784912109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.052520751953125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0455322265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0261993408203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0224456787109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.020477294921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01346588134765625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0129547119140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0081329345703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0081329345703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.18603515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1583251953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.119384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.08612060546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.103759765625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0950927734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0889892578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.06610107421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.059539794921875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.053558349609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.046142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.02691650390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.023040771484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0194091796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0150146484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0138092041015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01253509521484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0150146484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.167724609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1580810546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.14111328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.07843017578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0867919921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.080322265625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.07940673828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.071044921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0675048828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.04412841796875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.03839111328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.02203369140625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.019287109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0178070068359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01155853271484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0112762451171875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0077667236328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0077667236328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.229248046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2161865234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.193359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10723876953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11865234375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.10980224609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10858154296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.09716796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09246826171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06024169921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05230712890625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0300750732421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02606201171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0240020751953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.015228271484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00960540771484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2220458984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.19970703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1708984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.10107421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1175537109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.10821533203125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.104248046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0830078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.05975341796875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.0517578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.0299835205078125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0255126953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.02294921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0161895751953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01517486572265625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01158905029296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0161895751953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.086669921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0806884765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.070556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.040069580078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.045684814453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.042205810546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.04071044921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0357666015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.033905029296875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0231475830078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0201263427734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01157379150390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0099029541015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.00904083251953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.006069183349609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00574493408203125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0040740966796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0040740966796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.07293701171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.06787109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.059234619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.03369140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.038787841796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.035736083984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.03424072265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0301055908203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0286407470703125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0196533203125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.017059326171875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.00981903076171875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.00832366943359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.007598876953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.005123138427734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00482177734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.003406524658203125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.003406524658203125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1962890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.183349609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1622314453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09136962890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1026611328125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.09442138671875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.09259033203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0819091796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.07769775390625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.052093505859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0450439453125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.025970458984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02215576171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.020233154296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01336669921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0128173828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00807952880859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00807952880859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2098388671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1802978515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.141357421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09521484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.11700439453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1064453125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10028076171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0780029296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.07025146484375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06024169921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.052001953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0305938720703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0263671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0233154296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0175323486328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.015472412109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.16259765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.153076171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1368408203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.076171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.08441162109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.07818603515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.0771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.06903076171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.065673828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.042999267578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.03741455078125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.021484375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.018829345703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.01739501953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01129913330078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0110015869140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.007778167724609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.007778167724609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2281494140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.215087890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1920166015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10711669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11865234375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.10968017578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10833740234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.096923828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09222412109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.060333251953125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05242919921875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0301361083984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0264129638671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0243988037109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0158538818359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01544189453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01079559326171875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0158538818359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2174072265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.1949462890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.165771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.09912109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1162109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1060791015625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.10211181640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.085693359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.080810546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.05889892578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.050933837890625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.0297393798828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0255126953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.022979736328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0165252685546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0155487060546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01251983642578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0165252685546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.098876953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.09210205078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.08062744140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.04571533203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0521240234375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.048065185546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.046478271484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.040863037109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.038726806640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.026397705078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0229339599609375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0131988525390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0113067626953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0103302001953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.006927490234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.006572723388671875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00464630126953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.006927490234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.0809326171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.07550048828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.06597900390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0374755859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.04302978515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.039703369140625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.038055419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.033538818359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0318603515625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.021820068359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0189361572265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01090240478515625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0092620849609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.00846099853515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.005706787109375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.005390167236328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0037994384765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0037994384765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2115478515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.197998046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.175048828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09869384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1109619140625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10186767578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10003662109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0885009765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.083984375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.05633544921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.048583984375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.02813720703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02392578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0218505859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01448822021484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0138702392578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0086822509765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01448822021484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.20361328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.177001953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09423828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.10955810546875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10028076171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.09716796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07647705078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.06903076171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.056610107421875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.049041748046875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.02874755859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0254974365234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.02227783203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0165252685546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.015716552734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0142669677734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0165252685546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1697998046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.159912109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0797119140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.08837890625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08184814453125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08074951171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07208251953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.06866455078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.045166015625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.03924560546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0225677490234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0199737548828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0184478759765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01203155517578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01172637939453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0087432861328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01203155517578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2298583984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.21630859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1934814453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10760498046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11944580078125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1103515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10906982421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.097412109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09271240234375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.060699462890625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.052703857421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0302886962890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0262298583984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.024139404296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015777587890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.015350341796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00975799560546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015777587890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2293701171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2054443359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1748046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.10467529296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.12225341796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.112060546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.10784912109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0904541015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08502197265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.062164306640625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.053802490234375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.03125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0267333984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.02398681640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0171051025390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01274871826171875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1065673828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0994873046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.087158203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0494384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.05633544921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.052001953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.05023193359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.044281005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.042022705078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.028656005859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0249481201171875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0143585205078125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01251220703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01149749755859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00766754150390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00728607177734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005741119384765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00766754150390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.090576171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.08447265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.07403564453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.041961669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.048126220703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.04437255859375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.04266357421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.037628173828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0357666015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0244293212890625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0211944580078125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01221466064453125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.010406494140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.00952911376953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00640106201171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.006046295166015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004314422607421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00640106201171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.20947265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1959228515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1734619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1099853515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10101318359375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.09906005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.087646484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0831298828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.055877685546875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.04815673828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.02783203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0237579345703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.021697998046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0143890380859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01378631591796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00876617431640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0143890380859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2127685546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1820068359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09686279296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1173095703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.107177734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10076904296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07769775390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.06982421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.060516357421875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0526123046875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.030670166015625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02703857421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0236358642578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.017578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0162200927734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0160675048828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0162200927734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.172607421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.162353515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1448974609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0809326171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.08984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.083251953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0732421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.069580078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.045806884765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.03985595703125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0229034423828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.020233154296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0186920166015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01215362548828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01183319091796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00881195068359375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01215362548828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2320556640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.21875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1087646484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.12066650390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.111572265625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11016845703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.098388671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.093505859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.061279296875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.053253173828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.030609130859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0264892578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0243682861328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01593017578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0155029296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.009857177734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01593017578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.237060546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.21240234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1807861328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.10833740234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1260986328125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.115966796875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1116943359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09368896484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08819580078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06451416015625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.0557861328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.032440185546875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.02789306640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0250701904296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.017822265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0167694091796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.013641357421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0167694091796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.11346435546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.10589599609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0927734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.052642822265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.059783935546875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.05523681640625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.053466796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.047119140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.044647216796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.030364990234375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.026397705078125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0151824951171875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01312255859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01201629638671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00799560546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.007617950439453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005611419677734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00799560546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.092529296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.086181640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.075439453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0428466796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.049102783203125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.04534912109375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.043487548828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.038330078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.036346435546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0248565673828125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0216522216796875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0124359130859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01059722900390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.00969696044921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00652313232421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.006160736083984375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00438690185546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00652313232421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2117919921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.197998046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.174560546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0987548828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.11114501953125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1024169921875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10015869140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0885009765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.08380126953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.056396484375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.04888916015625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.02813720703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0240631103515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0219268798828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01453399658203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.013916015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00894927978515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01453399658203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.22705078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1986083984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1595458984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10516357421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.124755859375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11456298828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1087646484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.08673095703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.07843017578125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06439208984375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.055999755859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03265380859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.028717041015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.025390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0185546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0172576904296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.016387939453125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.016387939453125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.169189453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.158935546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1416015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.079345703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0882568359375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08160400390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08050537109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07171630859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.068115234375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0450439453125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0391845703125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0225372314453125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02001953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.01849365234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0120697021484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.011749267578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00902557373046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0120697021484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.231689453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2177734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1942138671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1083984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1202392578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11126708984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10986328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.09796142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09307861328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06103515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0531005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.030487060546875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0263671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0242156982421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0158233642578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.015380859375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.009735107421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0158233642578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2406005859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2161865234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.18408203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.11004638671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.127197265625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1173095703125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1134033203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09521484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0892333984375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06494140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.056304931640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.03271484375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0281524658203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.02532958984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0179595947265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0169677734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01352691650390625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01352691650390625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.11260986328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.10498046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.09185791015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.052276611328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.059478759765625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.05487060546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.052978515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.046722412109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.044281005859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0301513671875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0262298583984375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01508331298828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0129241943359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01183319091796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007904052734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.007495880126953125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0053558349609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007904052734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.0933837890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.087158203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.076171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.043304443359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.04962158203125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.045867919921875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.043975830078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.038787841796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.036773681640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0251312255859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0218963623046875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01256561279296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01068115234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0097808837890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0065765380859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.006198883056640625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004329681396484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0065765380859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.214111328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2001953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.17626953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09979248046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.112548828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10369873046875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1011962890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.08941650390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0845947265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.057098388671875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.049468994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0285186767578125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.024322509765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.022186279296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0147247314453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.014068603515625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0091094970703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0147247314453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.221435546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1943359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1568603515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10186767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1251220703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1143798828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10638427734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.08526611328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.07879638671875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06427001953125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.055633544921875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03240966796875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0274810791015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0244140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0179443359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01617431640625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01519012451171875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01617431640625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1707763671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1605224609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1427001953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.08026123046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.08929443359375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08251953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.081298828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0723876953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.06878662109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.045623779296875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0396728515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0228271484375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.020294189453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0187530517578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0122833251953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0119171142578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00930023193359375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0122833251953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.22998046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.216064453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1923828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10772705078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11968994140625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11065673828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10919189453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.09722900390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09222412109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06072998046875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.052825927734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03033447265625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0262451171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.024078369140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0157012939453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.015228271484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00980377197265625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0157012939453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.244384765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.21923828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.18603515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1121826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1300048828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.11956787109375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1153564453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09649658203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.09033203125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06646728515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.05755615234375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.033538818359375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.028594970703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.025634765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0185089111328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0174713134765625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0135955810546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0135955810546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.11962890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.111572265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.09765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.055450439453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06304931640625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0582275390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.056365966796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.049652099609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.046875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.031982421875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.027801513671875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0159759521484375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.013671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.012481689453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.008331298828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0079193115234375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0054931640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.008331298828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.0986328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.09210205078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0804443359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.045745849609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.05242919921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0484619140625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0465087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.040985107421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.03887939453125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0265350341796875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0231170654296875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01326751708984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01128387451171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01030731201171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.006931304931640625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.006542205810546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004547119140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.006931304931640625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.222900390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.20849609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1832275390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.11737060546875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.108154296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10552978515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.09320068359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.088134765625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.05963134765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05169677734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.029815673828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0254364013671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0232086181640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0155029296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0147857666015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0097503662109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0155029296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2320556640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2056884765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.1080322265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.127197265625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11651611328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11090087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.089599609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.08349609375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0653076171875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05670166015625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032928466796875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02899169921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0256195068359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01837158203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0171966552734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01593017578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01593017578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.174072265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.16357421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1455078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.081787109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.09124755859375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.0843505859375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.0828857421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07379150390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.07012939453125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.046630859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04052734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0233306884765625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.020782470703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.019195556640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0125732421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01221466064453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0096282958984375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0125732421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.23388671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2198486328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.19580078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1097412109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1219482421875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11279296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11126708984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.098876953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09393310546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06195068359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0538330078125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03094482421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.026702880859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.02447509765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01557159423828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0098724365234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.250244140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.224365234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1904296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1146240234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1324462890625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.121826171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1180419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09869384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.09228515625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06781005859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.05859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.034210205078125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0293121337890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0262908935546875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0189361572265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0179443359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01409912109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01409912109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1273193359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1185302734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10394287109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.059173583984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06707763671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06201171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.060089111328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05291748046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.05010986328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.034088134765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0296783447265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.017059326171875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01471710205078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01348876953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00896453857421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0085296630859375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0062713623046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0062713623046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1026611328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0958251953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.08355712890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.047607421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0548095703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.05072021484375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.048370361328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04266357421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.040496826171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0277557373046875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0242156982421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.013885498046875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01175689697265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0107574462890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007259368896484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.006832122802734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004802703857421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007259368896484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.231689453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2166748046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1907958984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10821533203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1220703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11236572265625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10980224609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.096923828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.09173583984375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.062042236328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0537109375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0309600830078125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.026458740234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.024169921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016082763671875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01535797119140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01012420654296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016082763671875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2379150390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2110595703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1695556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11090087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.12841796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11724853515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11383056640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.09112548828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.083251953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0662841796875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05718994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03363037109375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02984619140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.026153564453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0191192626953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.018310546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.016357421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.016357421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1710205078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.16064453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1429443359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.08056640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.089599609375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.0828857421875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08154296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.072509765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.06890869140625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.045745849609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.039886474609375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0229034423828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0203857421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.018829345703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.012298583984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01195526123046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00936126708984375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.012298583984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.236328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.221923828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.197509765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.11077880859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.12310791015625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11376953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1123046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.099853515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09478759765625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06256103515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.054351806640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0269775390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0247344970703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.016204833984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0157470703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.009979248046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.016204833984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.24462890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2196044921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.186279296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1119384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1304931640625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.11944580078125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1153564453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09649658203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0904541015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.0662841796875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.05731201171875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.0333251953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0284271240234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.025482177734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0181121826171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01702880859375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01329803466796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01329803466796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.12939453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.12054443359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10601806640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.060089111328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.068115234375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06292724609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.060943603515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0538330078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.051025390625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0345458984375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.030059814453125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0172576904296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.014801025390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0135498046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.009002685546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0085601806640625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00597381591796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00597381591796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1033935546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.09661865234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.08453369140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.048004150390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.055084228515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.050994873046875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.048675537109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0430908203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0408935546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0278778076171875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0243072509765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01392364501953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01180267333984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01081085205078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007259368896484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.006839752197265625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00470733642578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007259368896484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2359619140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.220947265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1947021484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11041259765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.114990234375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.09918212890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.093994140625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0634765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.054962158203125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.031707763671875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0269622802734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.024658203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016448974609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.015655517578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.010162353515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016448974609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.236083984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2147216796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1829833984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11004638671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1256103515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1151123046875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.09442138671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0880126953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06475830078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.056182861328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03302001953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02947998046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0265960693359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0190277099609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.018402099609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.016082763671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.016082763671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1766357421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.166015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.147705078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0831298828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.09228515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08551025390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08416748046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07489013671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.07110595703125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.04718017578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04107666015625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.023590087890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0208892822265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.019287109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01258087158203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0122222900390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.009368896484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01258087158203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2371826171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.222900390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1986083984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.111328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.12359619140625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1143798828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11279296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10040283203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0953369140625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06280517578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05462646484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0313720703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0271148681640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0248565673828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0162200927734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0157318115234375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.010040283203125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0162200927734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2471923828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2213134765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1871337890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.11279296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.13134765625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.12091064453125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11651611328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0970458984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0908203125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06683349609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.057769775390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.03363037109375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0284271240234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.02532958984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0182342529296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.017120361328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0128021240234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0128021240234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.12408447265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.11590576171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1019287109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.057708740234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06549072265625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.060546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.05859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05181884765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04913330078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.033233642578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0288848876953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0166168212890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01419830322265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0130157470703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00867462158203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0082550048828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00568389892578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0082550048828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1025390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.09588623046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0838623046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.047637939453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.05462646484375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.05059814453125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.048309326171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.042724609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.040557861328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.027618408203125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0241241455078125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0138092041015625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.011749267578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01076507568359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00720977783203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0067901611328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004734039306640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00720977783203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.24072265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.225830078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.19921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11279296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1273193359375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11724853515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1143798828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10137939453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.09613037109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.064697265625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0560302734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032318115234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.027557373046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0251922607421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0167694091796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0160064697265625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01039886474609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0167694091796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.224853515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.20361328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.16943359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10418701171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1258544921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11529541015625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.107666015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.09075927734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.08404541015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0648193359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.055694580078125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03265380859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0271148681640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0245208740234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.017822265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.016204833984375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.013580322265625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.016204833984375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.17529296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1646728515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.146728515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.08258056640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.09173583984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08489990234375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.0836181640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.074462890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0706787109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.046844482421875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04083251953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0208740234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0192718505859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0125579833984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01219940185546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00948333740234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0125579833984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2392578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.22509765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.200439453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.12481689453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11541748046875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11395263671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.101318359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09613037109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0633544921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05511474609375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03167724609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0274200439453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0251312255859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0164337158203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.015960693359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0102081298828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0164337158203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.24462890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2186279296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.184814453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.11181640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.13037109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1201171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1153564453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0960693359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.09002685546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06658935546875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.057525634765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.033538818359375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.028228759765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0251617431640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0182342529296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0170745849609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01285552978515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01285552978515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.129638671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.121337890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.106689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.060302734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.068359375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06317138671875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.061248779296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05426025390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.051513671875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03472900390625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0301666259765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0173797607421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01496124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01373291015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.009124755859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0087127685546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006221771240234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006221771240234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.10601806640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.09930419921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0869140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.04931640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.056365966796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.05224609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.050048828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0443115234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0419921875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0285491943359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0249176025390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01427459716796875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01215362548828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.011138916015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00745391845703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.007049560546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00489044189453125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00745391845703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.24609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2305908203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.204345703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.115234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1297607421875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1195068359375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.116943359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10369873046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.09844970703125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06585693359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.057098388671875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03289794921875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.028106689453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.025726318359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016998291015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0162811279296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0104827880859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0162811279296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.232421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2115478515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10736083984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.123291015625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.109619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.09136962890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0836181640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06341552734375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05474853515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032501220703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0291290283203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.02618408203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01898193359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0183868408203125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0163421630859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0163421630859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.17041015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.16015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0802001953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.08917236328125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08251953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.081298828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07232666015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0687255859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.04547119140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.039642333984375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0227813720703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0202178955078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.018646240234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01215362548828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0118255615234375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00905609130859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01215362548828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2347412109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2208251953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1968994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1102294921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1224365234375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11309814453125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1116943359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0994873046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09454345703125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.062255859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0540771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0311279296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0268707275390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.024658203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0162200927734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0157623291015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01001739501953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0162200927734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.237548828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.212158203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.179443359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1083984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.126708984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1165771484375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1119384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09307861328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08746337890625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.0645751953125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.05584716796875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.032501220703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.027374267578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0243988037109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0176239013671875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0164337158203125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01244354248046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0164337158203125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1268310546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1190185546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10479736328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.059112548828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.067138671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.061920166015625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.059967041015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.053192138671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.05059814453125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.034088134765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0295867919921875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0170440673828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.014556884765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01336669921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00888824462890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0084686279296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005828857421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005828857421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1043701171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0977783203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.08563232421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.04852294921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.05560302734375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.051483154296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.04925537109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04364013671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04150390625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0281829833984375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.02459716796875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01409149169921875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0119781494140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.010986328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0073699951171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00695037841796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0048675537109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0073699951171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.240966796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.22607421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.200927734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11285400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.126953125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11688232421875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11444091796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.1016845703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0965576171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06451416015625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.055816650390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03216552734375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.027496337890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.02520751953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016632080078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01593017578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01021575927734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016632080078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.226318359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.200927734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1597900390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10552978515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.12078857421875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11083984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10821533203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0863037109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0772705078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0621337890625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05377197265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03118896484375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0276336669921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.02398681640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.017181396484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.016448974609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01415252685546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.016448974609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1759033203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.165283203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1473388671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0826416015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0919189453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08502197265625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08380126953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0745849609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0709228515625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.046844482421875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.040771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0234832763671875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02069091796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.01910400390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01253509521484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01221466064453125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00901031494140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01253509521484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.23828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2242431640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.11181640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1241455078125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11474609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11334228515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.1009521484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09600830078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06304931640625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.054779052734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.031585693359375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0272979736328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.025054931640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0166015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0161285400390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.010284423828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0166015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2392578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2139892578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1806640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.109130859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.12744140625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.11761474609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11279296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09393310546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.088134765625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06500244140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.056549072265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.032623291015625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0277099609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0247802734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0173797607421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0160675048828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01285552978515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0160675048828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.127685546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1197509765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10577392578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.059539794921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0675048828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.062286376953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.060394287109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.053680419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.051116943359375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.034271240234375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.02978515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0171356201171875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.014678955078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0135040283203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00897979736328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00856781005859375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005931854248046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005931854248046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.107666015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1009521484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0888671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0501708984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.05731201171875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.052978515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.050872802734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.045196533203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04302978515625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0290374755859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.025299072265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.014495849609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0123443603515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0113372802734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007564544677734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.007160186767578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0048980712890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007564544677734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.244140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2294921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2039794921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11456298828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.128173828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1181640625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11602783203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.103271484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.09820556640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06512451171875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.056396484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032501220703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02777099609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0254669189453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016693115234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01003265380859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016693115234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2384033203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.213623046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1734619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11083984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.12646484375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11627197265625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.114013671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.09246826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0831298828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.065185546875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05645751953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032928466796875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02923583984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0257720947265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01849365234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01776123046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0152587890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0152587890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1817626953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1708984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1522216796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.08538818359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.09466552734375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.087646484375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08648681640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0732421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.04827880859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04205322265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.024139404296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0212860107421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0196380615234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01273345947265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0123748779296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0091094970703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01273345947265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2427978515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2283935546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.203857421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1138916015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.126220703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11688232421875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.115478515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10302734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09783935546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06414794921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05584716796875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.032012939453125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0277252197265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0254669189453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01654052734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.016082763671875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01019287109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01654052734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.240234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2142333984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1810302734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1094970703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1285400390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.11846923828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11322021484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0943603515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08856201171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.0653076171875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.05670166015625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.032684326171875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0276031494140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.024627685546875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.01739501953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01611328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.012451171875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01611328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.126220703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.11834716796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1044921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.058868408203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06695556640625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06170654296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0596923828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05303955078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.05059814453125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.034027099609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0294952392578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01702880859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01457977294921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0134124755859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.008941650390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00852203369140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005985260009765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005985260009765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.10760498046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.10089111328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.088623046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.050079345703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.057464599609375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.05316162109375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.050811767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04510498046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.042938232421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0291290283203125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.025360107421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01456451416015625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0123443603515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0113372802734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007610321044921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00717926025390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00496673583984375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007610321044921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.24365234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2288818359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.203369140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.1142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.128173828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1180419921875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11572265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10296630859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.09783935546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0650634765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.056304931640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032470703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0277557373046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.02545166015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0167236328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01015472412109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0167236328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2232666015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.19677734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1595458984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10345458984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.12030029296875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10968017578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1065673828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.08526611328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0772705078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0621337890625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.053375244140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.031524658203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0276031494140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0242919921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01806640625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0172271728515625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01496124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01496124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.185546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1744384765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1556396484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.08721923828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.096923828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.0894775390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08831787109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07879638671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.074951171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.049285888671875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.042999267578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.02471923828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0218353271484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0201873779296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.013214111328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01287841796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00955963134765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.013214111328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.242431640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.228271484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.2037353515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1138916015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.126220703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11676025390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1153564453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.1029052734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0977783203125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06414794921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0557861328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.032073974609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0277862548828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.025543212890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0167694091796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.016326904296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01047515869140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0167694091796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.238525390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2120361328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.178955078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1085205078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1279296875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1180419921875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11236572265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09332275390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0877685546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06500244140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.056427001953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.032501220703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.02716064453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0241546630859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0172119140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0158538818359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01186370849609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0158538818359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.121826171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1141357421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1007080078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.056671142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.064453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.059478759765625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.05743408203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05108642578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.048675537109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03265380859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.02838134765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.016326904296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.013916015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0128021240234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00853729248046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00812530517578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005523681640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00812530517578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.10430908203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0975341796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.08575439453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.04852294921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.055633544921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.051513671875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.04913330078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04364013671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04156494140625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0281829833984375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.024566650390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01409149169921875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.011932373046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0109710693359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00736236572265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.006938934326171875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004795074462890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00736236572265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.237060546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.22265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1978759765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11090087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.12451171875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11492919921875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11236572265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10003662109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.09515380859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06329345703125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0548095703125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03155517578125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0269622802734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0247344970703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0162811279296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0156097412109375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.009918212890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0162811279296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2266845703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1973876953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.15625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10498046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.12310791015625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1085205078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.085205078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0767822265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06365966796875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05517578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032379150390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.028167724609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0245513916015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0185089111328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.017425537109375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01538848876953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01538848876953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1859130859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1751708984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.15625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.08740234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0970458984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08856201171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0789794921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.07513427734375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.049468994140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04302978515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0247650146484375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02178955078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0201416015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0132293701171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01287078857421875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.009368896484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0132293701171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.239990234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2257080078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.2017822265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.112548828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1248779296875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1153564453125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11407470703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10174560546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09686279296875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06341552734375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.055145263671875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.031646728515625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.027496337890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0252838134765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0164794921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0160369873046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01041412353515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0164794921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2440185546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2169189453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.183349609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1112060546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.13134765625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.12060546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11505126953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09552001953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08990478515625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.0667724609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.0577392578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.033477783203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.028228759765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.025177001953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.01806640625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0167083740234375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01305389404296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0167083740234375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1246337890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.11676025390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1029052734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.058013916015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0660400390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06085205078125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.058837890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05224609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04974365234375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03350830078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0290374755859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0167694091796875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0142822265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0131378173828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00878143310546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0083770751953125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005733489990234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0083770751953125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1053466796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.09844970703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0863037109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.048858642578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.056365966796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.052154541015625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.04962158203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0440673828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.041961669921875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0285491943359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0248870849609375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0142974853515625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01207733154296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0111083984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007503509521484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.007061004638671875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0049591064453125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007503509521484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2451171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.23046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.204833984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11480712890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1285400390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11859130859375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1163330078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.103515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.098388671875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06536865234375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05657958984375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0325927734375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.027862548828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.025543212890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01678466796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0161285400390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01010894775390625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01678466796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2041015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1783447265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.143310546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0941162109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.108154296875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0994873046875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.09698486328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.076904296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0689697265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.055755615234375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.048492431640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.028228759765625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0253753662109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0222320556640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01534271240234375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.014068603515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.178466796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.168212890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1500244140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0838623046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.093017578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08612060546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.0849609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0758056640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0721435546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.04730224609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04119873046875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0236358642578125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0208282470703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0191802978515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01248931884765625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01215362548828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00870513916015625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01248931884765625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.232177734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.218505859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1087646484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1204833984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1114501953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11029052734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.098388671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0936279296875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.061187744140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.053192138671875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.030548095703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0264434814453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0243072509765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015777587890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.015350341796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00970458984375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015777587890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2314453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2054443359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.173583984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.10516357421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.12457275390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.11444091796875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.10894775390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0904541015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0853271484375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06341552734375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.0548095703125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.031829833984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.026611328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0237579345703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0172119140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0158843994140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01219940185546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0158843994140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1282958984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.12066650390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10693359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0599365234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0675048828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06256103515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06072998046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05413818359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.051544189453125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03424072265625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0298004150390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0171051025390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0146484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01346588134765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00890350341796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00852203369140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00560760498046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00560760498046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1126708984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.10601806640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.09375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0526123046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.059814453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.055206298828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.05328369140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04754638671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0452880859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0303192138671875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.026336669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01512908935546875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01287078857421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0118255615234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0078582763671875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.007480621337890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004924774169921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0078582763671875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2451171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2308349609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.205810546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.1148681640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.128173828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1182861328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11651611328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.1038818359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.098876953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06500244140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.056427001953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032470703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02789306640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0255889892578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0167083740234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01611328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01007080078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0167083740234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.22314453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.187255859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.138427734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.1024169921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1212158203125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11138916015625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1070556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07623291015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.07012939453125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06231689453125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05377197265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.031402587890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0266876220703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0217437744140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0173797607421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.016143798828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.013427734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.016143798828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.186767578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1756591796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1571044921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.08770751953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.09722900390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08990478515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.08880615234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07928466796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.07550048828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.049468994140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.043060302734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.02471923828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0217132568359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.020050048828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0130157470703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01267242431640625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0090179443359375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0130157470703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.23779296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.22412109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.200439453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.111572265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.123779296875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1143798828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1129150390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.1009521484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09625244140625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0628662109375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0545654296875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03143310546875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0272064208984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0250091552734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01641845703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0160064697265625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.010162353515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01641845703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2354736328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2093505859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1070556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.127197265625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.116943359375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11090087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09228515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0870361328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06451416015625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.055999755859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.032318115234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0270538330078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.024200439453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0172271728515625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.015777587890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01233673095703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.015777587890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.126953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1192626953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.059295654296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0670166015625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06195068359375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05352783203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0509033203125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03399658203125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.029632568359375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.016998291015625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0145416259765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01334381103515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0088653564453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.008453369140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005664825439453125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005664825439453125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.10845947265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1016845703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.08978271484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.050628662109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.05792236328125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.053466796875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.051300048828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.045654296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04351806640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0293121337890625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0255279541015625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01465606689453125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.012420654296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01141357421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007640838623046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00722503662109375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0048675537109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007640838623046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2447509765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.230224609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2052001953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11468505859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1282958984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1182861328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11627197265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10357666015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0986328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06500244140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.056365966796875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032470703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02783203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.025543212890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0167083740234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0160980224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01006317138671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0167083740234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.203369140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1771240234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1429443359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0938720703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1077880859375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0992431640625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.09710693359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07611083984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.06878662109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0555419921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.047943115234375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0279388427734375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0242919921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0208892822265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01537322998046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01461029052734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0120086669921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01537322998046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.19189453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1807861328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1614990234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.09014892578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.099853515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.09246826171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.09136962890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.08160400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.07757568359375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.050872802734375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04425048828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0254058837890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.022308349609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.02056884765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.013397216796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01305389404296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00921630859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.013397216796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.239501953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.225830078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.201904296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.11236572265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.12445068359375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11517333984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11383056640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.1016845703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09661865234375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.063232421875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.055023193359375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.031494140625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0272979736328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0251312255859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.016326904296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.015899658203125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01000213623046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.016326904296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.23779296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2120361328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1795654296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.108154296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1280517578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1177978515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09356689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08807373046875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06500244140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.056396484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.03277587890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.027252197265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0243988037109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.01788330078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.016571044921875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01222991943359375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.016571044921875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.128173828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.120361328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10650634765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.059844970703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06793212890625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06268310546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0606689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.054046630859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.051544189453125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.034454345703125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.029937744140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01727294921875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01470947265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.013519287109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0090484619140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00860595703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00577545166015625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00577545166015625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.111572265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.10479736328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0926513671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0521240234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.059417724609375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0548095703125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.05279541015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04705810546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04486083984375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03009033203125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0261688232421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0150299072265625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01274871826171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007808685302734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.007415771484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.004894256591796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.007808685302734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2437744140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2291259765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.204345703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11419677734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.127685546875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11785888671875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11578369140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10333251953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.09820556640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0648193359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05621337890625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0323486328125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0277557373046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0254364013671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0166168212890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0160369873046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01000213623046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0166168212890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2021484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.176025390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.135986328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.092041015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1075439453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0968017578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0948486328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.073486328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0660400390625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.054962158203125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0477294921875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.028656005859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0256805419921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0223846435546875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.017333984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0167999267578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0152587890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0167999267578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.197509765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1859130859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.166015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0928955078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.10284423828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.09521484375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.093994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.083984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.07989501953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.052398681640625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.045562744140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0261993408203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02294921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.021148681640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01377105712890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.013427734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.009368896484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01377105712890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.24365234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2293701171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.2049560546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.126953125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11712646484375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1158447265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.1033935546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0985107421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06451416015625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05596923828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.032257080078125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.027923583984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0257110595703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.016876220703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.016448974609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01056671142578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.016448974609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2371826171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2122802734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.18017578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1080322265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1275634765625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.11749267578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11181640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09356689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08831787109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06475830078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.056121826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.032440185546875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0269012451171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0240631103515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.01727294921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.015899658203125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0114593505859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.015899658203125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1309814453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1229248046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1090087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.061248779296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06927490234375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06396484375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06201171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.055206298828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0526123046875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03509521484375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0305328369140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.017547607421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01500701904296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01378631591796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0091400146484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00873565673828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005825042724609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005825042724609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.11334228515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1065673828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.09423828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.053009033203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.060455322265625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.055755615234375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.053741455078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.047821044921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.045654296875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0306549072265625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0266571044921875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01531219482421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01297760009765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.011932373046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0079498291015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00754547119140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00498199462890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0079498291015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.24951171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2347412109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2091064453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.1170654296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1307373046875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.12060546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11871337890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.105712890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.1005859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06640625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.057525634765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.033111572265625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0283966064453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0260467529296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016998291015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0164031982421875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01016998291015625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0164031982421875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.212890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1871337890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.148681640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0989990234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.11676025390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.107177734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1016845703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07965087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.07318115234375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.059967041015625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.051666259765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0301666259765625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.025360107421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.021575927734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016510009765625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0153350830078125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01201629638671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016510009765625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2021484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1903076171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.170166015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0950927734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.10552978515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.09747314453125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.0963134765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0819091796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0537109375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.046661376953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0268402099609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0235137939453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.02166748046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01416015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0138092041015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00959014892578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01416015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2449951171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.23095703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.2061767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1151123046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1275634765625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11798095703125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1165771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09893798828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06488037109375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05633544921875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03240966796875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0281219482421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.025848388671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0169677734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0165252685546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01058197021484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0165252685546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2392578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.214111328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.181640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.10906982421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.12841796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.11846923828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11273193359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09442138671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0889892578125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.065185546875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.0565185546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.0325927734375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0270233154296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.024139404296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0171966552734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.015777587890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0112152099609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.015777587890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1302490234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.12225341796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10833740234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06097412109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06915283203125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06378173828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06182861328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.054931640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.052459716796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03509521484375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.030487060546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0175323486328125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01493072509765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01371002197265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00913238525390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00870513916015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00572967529296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00572967529296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1148681640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.10784912109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.09539794921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0537109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06146240234375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.05645751953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.054412841796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04840087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04632568359375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0311279296875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0269622802734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0155792236328125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0131683349609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01207733154296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00811767578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00771331787109375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005069732666015625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00811767578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2509765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.23583984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.21044921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11785888671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.131591796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.12139892578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1195068359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10638427734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.10125732421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06683349609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.057952880859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03338623046875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0287017822265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0263519287109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0172271728515625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0166168212890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0106201171875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0166168212890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2164306640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1805419921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1412353515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0989990234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.11651611328125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10638427734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1031494140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07635498046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0692138671875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.05999755859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.052032470703125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0303802490234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0269317626953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0228271484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.017303466796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0162811279296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01519012451171875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0162811279296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.20458984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.192626953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.09619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.106689453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.09869384765625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.09747314453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0869140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0826416015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.054351806640625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.047271728515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.02716064453125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.023773193359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0219268798828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0143280029296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01396942138671875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00971221923828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0143280029296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.24609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.231689453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.20703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1156005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1280517578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11859130859375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1170654296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10455322265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09942626953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06512451171875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.056610107421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.032501220703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.028167724609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0258941650390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0169219970703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0164794921875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01039886474609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0164794921875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.23876953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2142333984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.181884765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1090087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.128662109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.11871337890625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.112548828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09478759765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08941650390625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.0655517578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.056854248046875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.032958984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.027435302734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0246124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0176849365234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.016265869140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01216888427734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.016265869140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.140869140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1322021484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06597900390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.07421875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06866455078125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.059356689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.05645751953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.037628173828125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.03277587890625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.018798828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.016082763671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.014739990234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0097198486328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00930023193359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006038665771484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006038665771484375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.12261962890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1151123046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10198974609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.05743408203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.065185546875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06005859375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.058319091796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0517578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04937744140625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0330810546875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.028656005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01654052734375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01406097412109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0128936767578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0085906982421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00820159912109375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005329132080078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00820159912109375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.25390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2386474609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.212890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11920166015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1329345703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.122802734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.12091064453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10760498046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.10235595703125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0675048828125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05853271484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03369140625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0289459228515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0265350341796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.017333984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.016754150390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01043701171875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.016754150390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.212646484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.189453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.15087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09906005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.11474609375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10601806640625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10125732421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.080810546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.073486328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.05889892578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.050811767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0294952392578125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.024810791015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0211639404296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0157318115234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01468658447265625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01071929931640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0157318115234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2081298828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.195556640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1748046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0980224609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.10870361328125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.10052490234375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.09930419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.08837890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0841064453125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.055328369140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.048095703125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0276641845703125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.024169921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0222625732421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01458740234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0142059326171875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0097503662109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01458740234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.24853515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2340087890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.208984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.11669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.12939453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1197509765625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.118408203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.100341796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06585693359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.057220458984375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.032806396484375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.028411865234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0261077880859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0170135498046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.016571044921875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0104522705078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.016571044921875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2445068359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2197265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1868896484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.11163330078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1309814453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.12078857421875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.115234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0970458984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.09136962890625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06634521484375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.057586669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.033172607421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0274810791015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.024566650390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.01739501953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.010986328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1392822265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1307373046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1156005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06536865234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0738525390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0682373046875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0662841796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.058837890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.055877685546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.037506103515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.032684326171875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0187530517578125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.016021728515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01470184326171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00977325439453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00933074951171875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00621795654296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00621795654296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.12164306640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.11407470703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.100830078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.056976318359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06488037109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.059722900390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0577392578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05126953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.048858642578125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03289794921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.028533935546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0164337158203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01393890380859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01277923583984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00853729248046875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00811004638671875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00531768798828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00811004638671875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.253173828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.238037109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2115478515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.118896484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1326904296875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.12237548828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.12060546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10711669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.101806640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0673828125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.058380126953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03363037109375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02880859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.026397705078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0172576904296875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0166473388671875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0102691650390625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0166473388671875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.21044921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.172607421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.12347412109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09039306640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.11431884765625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10260009765625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.09735107421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07037353515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.064208984375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.05743408203125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.050445556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.02947998046875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0257415771484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.022186279296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0170745849609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01548004150390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01568603515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01548004150390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.21337890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2008056640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1790771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10052490234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11181640625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.10321044921875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10186767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.09075927734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0863037109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0570068359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.049468994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.028656005859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.024993896484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.023040771484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01540374755859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01502227783203125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.010467529296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01540374755859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.25244140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.237548828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.2119140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.11871337890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1319580078125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.121826171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.12030029296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.107177734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.10186767578125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.067138671875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.058258056640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03363037109375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.029083251953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0267333984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0177764892578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.017333984375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01116180419921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01116180419921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.24462890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2208251953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.18798828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.112060546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1317138671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.12164306640625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11553955078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09783935546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.09222412109375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06683349609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.058197021484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.03350830078125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.027984619140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.025146484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.017669677734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0162200927734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01198577880859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0162200927734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1444091796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1353759765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1201171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0677490234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.076171875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0704345703125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06866455078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.061004638671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.057952880859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.038604736328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.033660888671875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.019287109375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.016510009765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01509857177734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00995635986328125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0095367431640625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00611114501953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00611114501953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.126953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.11920166015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1055908203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.05963134765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0673828125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06219482421875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0604248046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.053680419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.051055908203125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.034210205078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.029693603515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01708984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0145416259765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01332855224609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00885009765625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00844573974609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005443572998046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005443572998046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.259521484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2440185546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.21728515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.1221923828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.135986328125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.12548828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1236572265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10992431640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.1046142578125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0689697265625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.059906005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.034423828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.029571533203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0270843505859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0177001953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01708984375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0105438232421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0105438232421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2236328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1812744140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.14892578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10162353515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1209716796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11102294921875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10699462890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0784912109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0748291015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06207275390625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0535888671875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02618408203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0217437744140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0171661376953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0158538818359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01264190673828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0158538818359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2176513671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2047119140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1025390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11395263671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.105224609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1038818359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.09246826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.08795166015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.058135986328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.050384521484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.029083251953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0253753662109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.023345947265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015472412109375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01508331298828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01044464111328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.015472412109375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.255615234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2403564453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.2142333984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1202392578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1331787109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.12322998046875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.121826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.1083984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.10296630859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0677490234375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.058929443359375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.033843994140625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0292510986328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0268096923828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.017578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.017120361328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0107879638671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0107879638671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2486572265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.22412109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1905517578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.11376953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.13427734375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.12353515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.117431640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.099365234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0936279296875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06829833984375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.059051513671875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.034423828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.028472900390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.025604248046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.01861572265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.017181396484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01226806640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01226806640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1494140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.140380859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0701904296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0791015625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0731201171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.07122802734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.06317138671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.060028076171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.040191650390625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.034912109375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.02008056640625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.017181396484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0157470703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01042938232421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.009979248046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00653839111328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00653839111328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.131103515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.12298583984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10882568359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.061492919921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0697021484375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.064208984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.062408447265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.055389404296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.052734375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.035400390625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.030670166015625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0176849365234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01503753662109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01377105712890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00917816162109375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00876617431640625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00568389892578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00568389892578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.26123046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.245361328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.21826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.12286376953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.136962890625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1260986328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.12451171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.110595703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.1051025390625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0694580078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.060302734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.034637451171875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.02978515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0272369384765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0178070068359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.017242431640625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.010711669921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.010711669921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2349853515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.205810546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.160400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10845947265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.12646484375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11651611328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.112548828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0867919921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.07794189453125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0648193359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05596923828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.032440185546875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0273895263671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.02313232421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.017364501953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0161895751953125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01226806640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0161895751953125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.216064453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1810302734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10186767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11309814453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.10467529296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10321044921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.09185791015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.08734130859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0577392578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.050079345703125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.02886962890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0252227783203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.023193359375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01525115966796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01485443115234375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01031494140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01525115966796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.257080078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.24169921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.21533203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.12109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1343994140625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.122802734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.109130859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.10357666015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0684814453125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0594482421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.034210205078125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0296478271484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0271759033203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0179290771484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0174560546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0113067626953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0113067626953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2548828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.22998046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1961669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.11669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.13720703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.126708984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.120361328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.1019287109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.095947265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06964111328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.060546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.03485107421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.029083251953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0261077880859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0184326171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0169830322265625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01226806640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01226806640625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1455078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1365966796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.12078857421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.068359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0770263671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.07122802734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0693359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.061431884765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.05828857421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.039031982421875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.034027099609375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.019500732421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0166778564453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0152740478515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0101165771484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00966644287109375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006275177001953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006275177001953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1278076171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.119873046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10589599609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.05999755859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06793212890625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.062744140625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0609130859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05389404296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.05126953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03448486328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0299530029296875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0172271728515625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.014678955078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01343536376953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.008941650390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00853729248046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005580902099609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005580902099609375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.262451171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.24609375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2188720703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.12335205078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.137451171875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.126953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1251220703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.11090087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.1053466796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06976318359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.060546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03485107421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0299224853515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.027313232421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0178680419921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01727294921875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01070404052734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01070404052734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2125244140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.172119140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1336669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09674072265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1199951171875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.110107421875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07415771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.069580078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0616455078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.052886962890625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0308380126953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0246124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0202178955078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016387939453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0143585205078125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01129150390625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.016387939453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2159423828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2030029296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1805419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.101806640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11297607421875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1046142578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10333251953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0916748046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0870361328125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.057647705078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05010986328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0288543701171875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.025177001953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0231170654296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01520538330078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.014801025390625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.010223388671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01520538330078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.259033203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.24365234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.216796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.12200927734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.13525390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1253662109375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.12371826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10992431640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.10430908203125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06890869140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.059844970703125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03436279296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02972412109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0272064208984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0178375244140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0173187255859375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01092529296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01092529296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.26123046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.23583984375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.201171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.11968994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1412353515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1295166015625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1234130859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.1046142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.098388671875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.07135009765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.061920166015625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.035919189453125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.02984619140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0268707275390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0191650390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0177154541015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01276397705078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01276397705078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1495361328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1400146484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.12371826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0701904296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0791015625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.07318115234375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.07122802734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.06304931640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.05975341796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.040191650390625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.03497314453125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0200653076171875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0171661376953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0157012939453125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01038360595703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00994110107421875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0064544677734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0064544677734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1295166015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1212158203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10699462890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.060760498046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0689697265625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0635986328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.061676025390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.054534912109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0518798828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0350341796875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0303802490234375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01751708984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01486968994140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01361083984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0090789794921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0086517333984375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00565338134765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00565338134765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.269287109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.25244140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2242431640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.126708984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1412353515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.13037109375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1285400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.11376953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.10791015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.07177734375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.062286376953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.035797119140625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.03076171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.028045654296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.018402099609375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0177764892578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01097869873046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01097869873046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.220703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.179443359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1298828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.097900390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.12115478515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11114501953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07330322265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0672607421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.062042236328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.05328369140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0310516357421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.024658203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0198516845703125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01654052734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.014556884765625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0108642578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01654052734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.21435546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.201416015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.17919921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10113525390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11236572265625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.10400390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10272216796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.091064453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.08648681640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.057281494140625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.049835205078125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.028656005859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02508544921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.023040771484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0151519775390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01473236083984375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0103607177734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0151519775390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2587890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.242919921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.21630859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1219482421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.13525390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1251220703125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1236572265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10968017578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.1041259765625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0689697265625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05987548828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03436279296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02978515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.02728271484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0178985595703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0173797607421875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.011199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.011199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.260498046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2353515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.200439453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1195068359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.140380859375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.129150390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.12310791015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.1043701171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0982666015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.07135009765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.061767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.035919189453125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.029693603515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.026702880859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0193328857421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0178985595703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01247406005859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01247406005859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.146240234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.136962890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.120849609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06866455078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.07769775390625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.07177734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0697021484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.06158447265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0584716796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0394287109375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0343017578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01971435546875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.016815185546875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01535797119140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01021575927734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00974273681640625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006378173828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006378173828125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1280517578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10577392578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06011962890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06817626953125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06298828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06103515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05401611328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.051239013671875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03460693359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0300750732421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0173187255859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01474761962890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01345062255859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00896453857421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00855255126953125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00560760498046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00560760498046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.26513671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2486572265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.220703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.12481689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1392822265625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1285400390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.12646484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.11212158203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.1063232421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0706787109375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.061370849609375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0352783203125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0302734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.027618408203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0180816650390625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0174713134765625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0107421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0107421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.2197265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.185302734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1494140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.10003662109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.11944580078125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10858154296875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.10443115234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07916259765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.07373046875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06097412109375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.052398681640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03082275390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0259246826171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0218658447265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0172119140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01279449462890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01605224609375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2127685546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.19970703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1776123046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.100341796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.111572265625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1031494140625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10186767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.09014892578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.08563232421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.05694580078125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0494384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.028472900390625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0249176025390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0228271484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01507568359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01468658447265625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01029205322265625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01507568359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2587890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.242919921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.216064453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.12200927734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1356201171875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.125244140625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.12384033203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10968017578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.10406494140625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06903076171875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05999755859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0345458984375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0299072265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.027374267578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.018157958984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.017669677734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01148223876953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01148223876953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.25927734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2342529296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1990966796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.11895751953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1412353515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1297607421875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.12274169921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.10400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.097900390625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.07171630859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.06207275390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.036224365234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.02972412109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0267333984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0196533203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.018096923828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01271820068359375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01271820068359375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1494140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.139892578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.123291015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.07025146484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.079345703125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0733642578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0712890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0628662109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.059600830078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.040252685546875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.03509521484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0200958251953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.017181396484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01568603515625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01043701171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0099639892578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006565093994140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006565093994140625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1280517578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1199951171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10540771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0601806640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0684814453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06317138671875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06109619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.05389404296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.05120849609375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.034759521484375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0301971435546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0173797607421875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0147705078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01348876953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00905609130859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00860595703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0057220458984375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0057220458984375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.269287109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.252197265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2235107421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.1268310546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1412353515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1304931640625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.1285400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.11370849609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.107666015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.07177734375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.06231689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03582763671875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.03076171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.028045654296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0184173583984375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01776123046875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.011016845703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.011016845703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.197265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1634521484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1234130859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.08880615234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.10968017578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.10125732421875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.093994140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0684814453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.06390380859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.05615234375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0484619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.028076171875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.022430419921875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0184478759765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.014862060546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01302337646484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0100555419921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.014862060546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2125244140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.19921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1771240234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1002197265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11151123046875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.10302734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10162353515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.090087890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.08538818359375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.056854248046875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04937744140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0284271240234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.024871826171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.022796630859375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01507568359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01464080810546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01024627685546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01507568359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.252685546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2369384765625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.2105712890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.11895751953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.13232421875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.12237548828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.12078857421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.1070556640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.1014404296875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.06744384765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.05853271484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.033721923828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0291595458984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0266265869140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.017669677734375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.017181396484375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01103973388671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01103973388671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.258544921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.23291015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.197509765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.118408203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.14013671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1295166015625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.12249755859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.1033935546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0972900390625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.0712890625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.06182861328125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.035675048828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0295562744140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0265655517578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0188446044921875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0172271728515625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.012603759765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.012603759765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1451416015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1357421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.11956787109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06805419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.07696533203125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.07122802734375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.069091796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.06097412109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0577392578125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.039093017578125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.034027099609375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0195465087890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.016693115234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01522064208984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0101318359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00966644287109375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006359100341796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.006359100341796875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.12420654296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.11614990234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1021728515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.05828857421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0662841796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.061187744140625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0592041015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.052215576171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.049560546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03363037109375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.029205322265625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.016815185546875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.014312744140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01308441162109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00876617431640625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.008331298828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005535125732421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.008331298828125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.268798828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.251953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.22314453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.12646484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1409912109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.13037109375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.12841796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.11358642578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.10748291015625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0716552734375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.062286376953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0357666015625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0307464599609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.02801513671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01837158203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0177459716796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01097869873046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01097869873046875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.203369140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.170654296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1295166015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09326171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1083984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.09967041015625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.09722900390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07000732421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.064208984375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0555419921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0478515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.02777099609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0235748291015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0189971923828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01490020751953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01398468017578125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01053619384765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01490020751953125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2103271484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1973876953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1754150390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.09942626953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.11029052734375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1021728515625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.10076904296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0892333984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0845947265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.05621337890625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.048919677734375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.028106689453125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.024627685546875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.022552490234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01476287841796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01433563232421875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01007080078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01476287841796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.24658203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2313232421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.2056884765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.116455078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1292724609375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.11956787109375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11810302734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.10443115234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0989990234375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0657958984375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.0572509765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0328369140625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0284271240234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0259552001953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01702880859375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0165557861328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01059722900390625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.0165557861328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.258056640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.230712890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1177978515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.1402587890625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1292724609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1220703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.10247802734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.096435546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.07110595703125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.061767578125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.03564453125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0292816162109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0262603759765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.018829345703125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.017120361328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01233673095703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01233673095703125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.140380859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.131103515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.1151123046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06591796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.07476806640625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06927490234375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.0589599609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.05584716796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0379638671875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0330810546875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0189971923828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0161590576171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.014739990234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00988006591796875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0093994140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0062255859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0062255859375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.11749267578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.10968017578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.09619140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0550537109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.06280517578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.05810546875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.055877685546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.049285888671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.046661376953125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0318603515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.027740478515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01593017578125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01352691650390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0123291015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00830841064453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0078582763671875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005260467529296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00830841064453125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.26416015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.24755859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2191162109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.124267578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1387939453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1282958984375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.126220703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.11138916015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.1055908203125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.070556640625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.06121826171875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.03521728515625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0302734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.027557373046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0181121826171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0174713134765625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.010955810546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.010955810546875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.208251953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1707763671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.12890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.09405517578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1124267578125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.103271484375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0997314453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.07159423828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.064697265625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0576171875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.049713134765625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0289764404296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0241241455078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01983642578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01568603515625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01435089111328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01154327392578125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01568603515625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2017822265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1890869140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.167724609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.09521484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.10601806640625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.097900390625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.09661865234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.08544921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0809326171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.05401611328125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04693603515625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0270233154296875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0236358642578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.021636962890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01422882080078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.013824462890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.009735107421875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01422882080078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2342529296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.2196044921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.1104736328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.12274169921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.1136474609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.11212158203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.09912109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.09393310546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.054290771484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.03125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02703857421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.02471923828125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0162506103515625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01580810546875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01021575927734375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.0162506103515625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.250244140625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.22314453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1885986328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.1143798828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.136962890625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1253662109375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.11865234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09912109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.09326171875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.06951904296875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.06005859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.034942626953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.02886962890625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0258636474609375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0187835693359375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01715087890625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01294708251953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01294708251953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.132080078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.12335205078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10809326171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06195068359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0706787109375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06524658203125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06298828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.055419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.052490234375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03594970703125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0312347412109375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0179901123046875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0153045654296875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01397705078125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00943756103515625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00894927978515625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00615692138671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00615692138671875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1138916015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1063232421875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.093017578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0533447265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.061279296875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.056549072265625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.054229736328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04766845703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04522705078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0311431884765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.02703857421875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0155792236328125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01317596435546875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01201629638671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0081634521484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.007701873779296875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00525665283203125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0081634521484375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.253662109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.2371826171875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.209716796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.119384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1336669921875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.123291015625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.12127685546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.1068115234375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.10107421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06793212890625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.058929443359375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.033935546875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0290679931640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0264892578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.017486572265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0168609619140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01067352294921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01067352294921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.132080078125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.11114501953125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.088623046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.05908203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0712890625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06512451171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.06268310546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04730224609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0438232421875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.03656005859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.032073974609375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.018646240234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0167694091796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01476287841796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.010772705078125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0099945068359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01035308837890625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0167694091796875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1904296875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.178466796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.158447265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0899658203125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0999755859375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.09246826171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.09124755859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0806884765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.07647705078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.051177978515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.04437255859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0255889892578125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0224151611328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0205535888671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01361846923828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.013214111328125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00946044921875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01361846923828125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.2166748046875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.180419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.10235595703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.1138916015625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.105224609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.1038818359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.091796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.08685302734375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.058074951171875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.050384521484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0290374755859375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.0251617431640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.02301025390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01531219482421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01486968994140625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0098876953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01531219482421875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.2384033203125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.2109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1778564453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.10858154296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.130126953125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.1192626953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.1131591796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.09368896484375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.08819580078125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.0665283203125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.05731201171875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.033538818359375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.02789306640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0250091552734375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.0183868408203125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.016876220703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0133819580078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.0133819580078125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.12939453125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.12103271484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.10614013671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.060791015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.069580078125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06402587890625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.061737060546875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.054351806640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.051727294921875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.035369873046875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0306396484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.017730712890625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0150146484375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01371002197265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.00931549072265625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00884246826171875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005992889404296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005992889404296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1099853515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.1026611328125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.08990478515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.051605224609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.05926513671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.054718017578125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0523681640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.046142578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04376220703125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.030120849609375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0261383056640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0150909423828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0127410888671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.01163482666015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0079193115234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0074615478515625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.005115509033203125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0079193115234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.257568359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.240966796875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2132568359375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.121337890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.1356201171875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.1251220703125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.123291015625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.108642578125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.10284423828125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0689697265625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.059722900390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.034423828125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.029571533203125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0269317626953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0177154541015625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.01708984375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0108489990234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0108489990234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.1492919921875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.12054443359375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.08990478515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.06805419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.07965087890625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.0726318359375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0712890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.04815673828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.04510498046875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.041015625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.035400390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0208740234375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0184173583984375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0148468017578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.01192474365234375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0113372802734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01025390625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0148468017578125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.157470703125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1473388671875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.130859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.07464599609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.08319091796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.07684326171875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.07586669921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.067138671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.06365966796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.04296875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.037628173828125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0217437744140625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.020263671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.01885986328125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01242828369140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.012115478515625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.01143646240234375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01242828369140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1749267578125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1636962890625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1456298828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.0831298828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0927734375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.08538818359375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.084228515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.07470703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.07086181640625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.047760009765625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.041778564453125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.024261474609375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.02252197265625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.020965576171875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01395416259765625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01363372802734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.0126495361328125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01395416259765625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.1878662109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.1640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.1376953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.08544921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.10528564453125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.09417724609375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.08905029296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.0732421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.0693359375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.053680419921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.045989990234375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.027557373046875, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0233306884765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.0211181640625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.016265869140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.01491546630859375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.01320648193359375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.016265869140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.q_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.12164306640625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.11346435546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.0994873046875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.05706787109375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.0657958984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06024169921875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.057891845703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.050933837890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.048553466796875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.033447265625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0288238525390625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0167388916015625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0140838623046875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.012847900390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0087890625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00832366943359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00565338134765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.00832366943359375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.k_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.0972900390625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.0909423828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.07940673828125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.0455322265625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.052734375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.048431396484375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.046234130859375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.040802001953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0386962890625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.02679443359375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0232086181640625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.01343536376953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01129150390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.0103302001953125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0070953369140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.006656646728515625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.00464630126953125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0070953369140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.v_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.243896484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.22802734375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.2020263671875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.11505126953125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.130126953125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.11871337890625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.11676025390625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.10302734375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.0980224609375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.06646728515625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.0570068359375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.0335693359375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.0284271240234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.02593994140625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.017608642578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.0169677734375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01113128662109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.01113128662109375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.o_proj", + "numel": 26214400, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17529296875, + "total_bits": 57024000.0, + "err": 0.128173828125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 62266880.00000001, + "err": 0.096435546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72529296875, + "total_bits": 71441920.0, + "err": 0.07257080078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22529296875, + "total_bits": 84549120.0, + "err": 0.056182861328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 79464320.0, + "err": 0.07763671875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 81927680.0, + "err": 0.06378173828125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17529296875, + "total_bits": 83238400.0, + "err": 0.0606689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.52529296875, + "total_bits": 92413440.0, + "err": 0.041839599609375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.662646484375, + "total_bits": 96014080.0, + "err": 0.040283203125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 105678720.0, + "err": 0.0389404296875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 108142080.0, + "err": 0.03265380859375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 131893120.0, + "err": 0.021148681640625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22529296875, + "total_bits": 136977920.0, + "err": 0.01849365234375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.016357421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 158107520.0, + "err": 0.0138702392578125, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2313232421875, + "total_bits": 163350400.0, + "err": 0.012969970703125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 212999679.99999997, + "err": 0.0135650634765625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.52529296875, + "total_bits": 144842240.0, + "err": 0.016357421875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.gate_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.1412353515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1322021484375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.1171875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.06646728515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0740966796875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.06854248046875, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.06744384765625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.0595703125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.0565185546875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.037811279296875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.032806396484375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.0189056396484375, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.016510009765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.01513671875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.00998687744140625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.00968170166015625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.006900787353515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.006900787353515625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.up_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1751085069444445, + "total_bits": 153951744.0, + "err": 0.13916015625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3751085069444446, + "total_bits": 168107520.0, + "err": 0.1302490234375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7251085069444443, + "total_bits": 192880128.0, + "err": 0.115478515625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2251085069444443, + "total_bits": 228269568.0, + "err": 0.06591796875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.031277126736111, + "total_bits": 214550400.0, + "err": 0.0736083984375, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1251085069444446, + "total_bits": 221191680.0, + "err": 0.06787109375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1751085069444445, + "total_bits": 224730624.0, + "err": 0.06689453125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5251085069444446, + "total_bits": 249503232.0, + "err": 0.059173583984375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6625542534722224, + "total_bits": 259231488.0, + "err": 0.056243896484375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031277126736111, + "total_bits": 285329280.0, + "err": 0.0380859375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125108506944445, + "total_bits": 291970560.0, + "err": 0.032989501953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031277126736111, + "total_bits": 356108160.0, + "err": 0.019195556640625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.225108506944444, + "total_bits": 369827328.0, + "err": 0.017425537109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0161285400390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031277126736111, + "total_bits": 426887040.0, + "err": 0.01079559326171875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.231277126736111, + "total_bits": 441042816.0, + "err": 0.01052093505859375, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125108506944445, + "total_bits": 575086080.0, + "err": 0.00914764404296875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.525108506944444, + "total_bits": 391060992.0, + "err": 0.0161285400390625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.down_proj", + "numel": 70778880, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1715892650462965, + "total_bits": 153702656.0, + "err": 0.0943603515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37529296875, + "total_bits": 168120576.0, + "err": 0.08209228515625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.727144820601852, + "total_bits": 193024256.0, + "err": 0.06793212890625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.227144820601852, + "total_bits": 228413696.0, + "err": 0.04296875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0313232421875, + "total_bits": 214553664.0, + "err": 0.055084228515625, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12529296875, + "total_bits": 221204736.0, + "err": 0.048614501953125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1715892650462965, + "total_bits": 224481536.0, + "err": 0.044921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.523441116898148, + "total_bits": 249385216.0, + "err": 0.03680419921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6644983362268517, + "total_bits": 259369088.0, + "err": 0.03515625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0313232421875, + "total_bits": 285332544.0, + "err": 0.028289794921875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12529296875, + "total_bits": 291983616.0, + "err": 0.024078369140625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0313232421875, + "total_bits": 356111424.0, + "err": 0.01468658447265625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.227144820601852, + "total_bits": 369971456.0, + "err": 0.0125885009765625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.523441116898148, + "total_bits": 390942976.0, + "err": 0.011566162109375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0313232421875, + "total_bits": 426890304.0, + "err": 0.00878143310546875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0079803466796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12529296875, + "total_bits": 575099136.0, + "err": 0.00818634033203125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.235026945891204, + "total_bits": 441308224.0, + "err": 0.0079803466796875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + } + ], + "last_module_idx": 82, + "base_perplexity": 5.979936015162279 +} \ No newline at end of file