{ "measurement": [ { "key": "model.layers.0.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.0515446811914444, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.051185011863708496, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.026728002354502678, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.026595551520586014, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.02650294452905655, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.01709524542093277, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0712495893239975, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05077889934182167, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.02651645988225937, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.026442455127835274, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.02613804303109646, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.028380703181028366, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026413120329380035, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01912635564804077, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.01704813353717327, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.018293945118784904, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.017033999785780907, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.016335979104042053, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.017031703144311905, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.016333987936377525, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015961075201630592, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017031151801347733, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015153263695538044, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.016332529485225677, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.0515446811914444, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.0515446811914444, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.0.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.051888685673475266, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.051526572555303574, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.022455183789134026, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.022337667644023895, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.022252174094319344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.012493480928242207, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.06880488991737366, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.05109550803899765, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.022266235202550888, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.022196440026164055, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.022422093898057938, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.023228628560900688, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.022173848003149033, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.014676572754979134, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.012448538094758987, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.014703407883644104, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.01243417989462614, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.0116649828851223, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.012432001531124115, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.011663195677101612, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.012290704064071178, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.01243174634873867, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.011411914601922035, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.01166163943707943, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.051888685673475266, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.051888685673475266, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.0.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.16849948465824127, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.1422615498304367, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.11506197601556778, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.09410161525011063, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.07904711365699768, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.05728914216160774, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.15315325558185577, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.10827884823083878, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.08175335824489594, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.06291574239730835, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.07022751867771149, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.08797664940357208, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.054069794714450836, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.038518644869327545, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.03366677835583687, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.048474155366420746, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.019955163821578026, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.016969285905361176, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.0167371928691864, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.01340905949473381, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.026044780388474464, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.015083486214280128, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01129976473748684, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.00874419417232275, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.11506197601556778, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, "best_option": { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.11506197601556778, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.0.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.018048683181405067, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.01263423077762127, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.009894996881484985, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.009090469218790531, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.007928614504635334, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.005497607868164778, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.011124419048428535, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.009855587966740131, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.008520212024450302, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.00557005824521184, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.005572076886892319, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.005667088087648153, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.0049598971381783485, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.004132694099098444, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.003918378613889217, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.002974356524646282, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.00262974901124835, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0024836582597345114, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.002298744861036539, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.00218260963447392, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0018380561377853155, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0022267652675509453, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.001599636161699891, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0019581085070967674, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.018048683181405067, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.018048683181405067, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.15095368027687073, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.14233887195587158, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.13947735726833344, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.12643775343894958, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.06994408369064331, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.06724482029676437, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.07740512490272522, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.0716518983244896, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07049666345119476, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0631726086139679, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.060316190123558044, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.03931840509176254, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03426358848810196, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03353874385356903, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.03337349742650986, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.01965194195508957, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.017588268965482712, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.01741449534893036, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01628781482577324, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.016177503392100334, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.010559458285570145, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.011099670082330704, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.010328729636967182, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00800111424177885, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.07740512490272522, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.07740512490272522, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2135459929704666, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.20192590355873108, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1982744038105011, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18022024631500244, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0993751659989357, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09571931511163712, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10960973054170609, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10138368606567383, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10010822117328644, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08997305482625961, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08591287583112717, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.055514488369226456, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0482499934732914, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0473635196685791, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.047145843505859375, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02768981084227562, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02414979599416256, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02392108365893364, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.022250521928071976, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.0221165232360363, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014405135996639729, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014213851653039455, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.0141018470749259, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009066724218428135, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10960973054170609, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10960973054170609, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.09874424338340759, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.08902619034051895, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.08506868034601212, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.07604729384183884, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.043627675622701645, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.040450289845466614, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.05155382677912712, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.04734690114855766, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.04497113823890686, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.038712598383426666, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.03639267385005951, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.026078026741743088, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.022673513740301132, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.02102668769657612, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.020622676238417625, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.013113741762936115, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.011357264593243599, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.011256631463766098, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.010464669205248356, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.010187416337430477, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.007228120695799589, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.00778307905420661, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.006716117728501558, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.005920155439525843, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.09874424338340759, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.09874424338340759, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.1.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.01835786923766136, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.01811937801539898, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.009405337274074554, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.009339241310954094, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.009301544167101383, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.00638729240745306, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.026573050767183304, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.017969967797398567, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.009323090314865112, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.009239871986210346, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.009586967527866364, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.01068209856748581, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.009225845336914062, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.006978969555348158, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.006342307664453983, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.006885370705276728, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.0063219801522791386, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0061225490644574165, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.0063178567215800285, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.006118907127529383, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.006107494235038757, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0063170683570206165, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.005860127974301577, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006116658914834261, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.01835786923766136, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.01835786923766136, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.1.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.012739398516714573, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.012531104497611523, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.006143445614725351, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.0060739414766430855, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.006033618468791246, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.0029974456410855055, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.017146576195955276, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.01240604929625988, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.006055758334696293, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.005971057340502739, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.006155670620501041, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.006752482615411282, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.005957664456218481, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.00361481960862875, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.002934891963377595, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.0036319296341389418, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.00290306587703526, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.0026535578072071075, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.002897075843065977, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.0026479451917111874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.002813484985381365, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.002895891899242997, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.002529246499761939, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.002644306281581521, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.012739398516714573, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.012739398516714573, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.1.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.13531209528446198, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.11281485110521317, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.09222806990146637, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07550786435604095, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.061972811818122864, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.045931532979011536, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11655787378549576, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.08375317603349686, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.06411343812942505, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.04854310303926468, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.05299799516797066, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06566964834928513, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.04104986786842346, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.030116155743598938, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.026953695341944695, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03529965877532959, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.015827376395463943, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.013632371090352535, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.013230626471340656, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.010853176936507225, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.018161455169320107, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.011838149279356003, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.008739654906094074, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.00705553125590086, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.11281485110521317, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, "best_option": { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.11281485110521317, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.1.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.05416184663772583, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.036251213401556015, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.02612174116075039, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.024866091087460518, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.02248549647629261, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.014182571321725845, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.03354430943727493, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.030307766050100327, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.025101078674197197, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.016252564266324043, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.016641635447740555, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.017099758610129356, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.015150250867009163, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.011886532418429852, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.011000091210007668, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.008992287330329418, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.00777824642136693, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.007195545360445976, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.006943603046238422, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.006497592199593782, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.005575584247708321, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0068177999928593636, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0047263833694159985, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006026421673595905, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.05416184663772583, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.05416184663772583, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.150831937789917, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.14197812974452972, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.13905881345272064, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.12602096796035767, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.06979519128799438, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.06705351918935776, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.07708780467510223, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.07149696350097656, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07038445770740509, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.06299693137407303, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.06014842540025711, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.03934529796242714, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.034501250833272934, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.033774103969335556, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.0336010567843914, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.019740788266062737, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.01819687895476818, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.018035918474197388, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.016942188143730164, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.016844922676682472, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.010884000919759274, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012129711918532848, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01066057849675417, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009398387745022774, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.07708780467510223, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.07708780467510223, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.18272539973258972, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17239397764205933, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16912266612052917, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1535940170288086, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0845523551106453, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08130909502506256, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09342020004987717, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08639504760503769, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08518323302268982, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07639200985431671, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07299961894750595, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04729386419057846, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.041100747883319855, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0402921661734581, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.040100980550050735, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02361414209008217, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.020651910454034805, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02044605277478695, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01903344690799713, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.018912047147750854, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01241605170071125, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012333326041698456, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012150314636528492, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00809439830482006, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09342020004987717, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09342020004987717, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.058902300894260406, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.049545954912900925, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.03442040830850601, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.030982067808508873, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.024352187290787697, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.017756901681423187, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.04367072507739067, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.0413096658885479, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.026103610172867775, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.020662296563386917, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.01952372118830681, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.017983196303248405, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.01769297569990158, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.012254602275788784, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.01184793096035719, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.010409963317215443, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.008898154832422733, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.006399729289114475, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.008415237069129944, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.005507756024599075, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.007839829660952091, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.008061439730226994, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.00439589936286211, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.004516853950917721, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.058902300894260406, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.058902300894260406, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.2.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.02984984964132309, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.025242241099476814, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.019668690860271454, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.017578501254320145, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.013558992184698582, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.009593999944627285, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.02192980796098709, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.01956290192902088, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.014180568046867847, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.011218119412660599, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.011370845139026642, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.011214778758585453, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.009434392675757408, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.006781619507819414, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.006004689261317253, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.005658514332026243, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.0039010210894048214, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0034614233300089836, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.0034960980992764235, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.0030423346906900406, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.003046007826924324, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.003219226375222206, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0021775350905954838, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.002436487702652812, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.02984984964132309, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.02984984964132309, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.2.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.02951953373849392, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.02519272081553936, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.018371814861893654, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.01643170788884163, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.013059472665190697, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.008621600456535816, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.02254847064614296, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.02052358351647854, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.013712120242416859, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.011122506111860275, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.011575885117053986, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.01155179738998413, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.009754336439073086, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.00641573965549469, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.0053571914322674274, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.00577600346878171, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.0034970156848430634, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.002888739574700594, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.0031402125023305416, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.002492207568138838, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.0030198986642062664, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.002896455582231283, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.0018648435361683369, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0019028203096240759, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.02951953373849392, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.02951953373849392, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.2.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.17525456845760345, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.15390843152999878, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.14423319697380066, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.12506738305091858, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.08111219108104706, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.071752168238163, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.09858648478984833, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.08991726487874985, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.08313003927469254, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.06664855778217316, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.06274663656949997, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.05032472312450409, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.0431237667798996, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.03897349163889885, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.03792514652013779, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.025203196331858635, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.019805127754807472, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.018940795212984085, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.01683189533650875, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.016099460422992706, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.012864596210420132, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.012284788303077221, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.011213739402592182, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.007347128354012966, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.09858648478984833, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.09858648478984833, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2237592339515686, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.16963566839694977, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.14761553704738617, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12565982341766357, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.1016063541173935, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07907918095588684, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12627306580543518, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11527271568775177, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1060803234577179, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06986598670482635, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06778624653816223, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0652254968881607, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.056606948375701904, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05046500265598297, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04894671216607094, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.033419739454984665, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02867545560002327, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027262914925813675, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.023353485390543938, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02234988659620285, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0194539837539196, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.021435711532831192, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017677193507552147, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.017462803050875664, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11527271568775177, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11527271568775177, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.19754347205162048, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1861264854669571, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18245258927345276, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16554643213748932, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0912209302186966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.0876825749874115, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10099603235721588, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09329365938901901, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.0919729471206665, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08229773491621017, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07869166135787964, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.051294948905706406, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04453928768634796, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.043641552329063416, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04342043772339821, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.025654075667262077, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.022760940715670586, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02254117839038372, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.021034322679042816, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.020894287154078484, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013775684870779514, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014172355644404888, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013486603274941444, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010013602674007416, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10099603235721588, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10099603235721588, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2341393530368805, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22117066383361816, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2170535773038864, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19728988409042358, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10871733725070953, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10456158965826035, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12004958838224411, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11103510856628418, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10954701155424118, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09821493178606033, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09386254101991653, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06076890602707863, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05279654636979103, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.051759667694568634, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05151572823524475, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030318357050418854, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026344968006014824, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026070713996887207, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024218536913394928, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024061134085059166, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01575702615082264, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015427102334797382, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015413369052112103, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009712770581245422, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12004958838224411, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12004958838224411, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.209426611661911, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.19013391435146332, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.18292976915836334, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.16519246995449066, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.09524887800216675, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.08871204406023026, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.11140166968107224, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.10189669579267502, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.09799934923648834, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.08475635200738907, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08014881610870361, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.0564190037548542, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.04897082597017288, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.04593935236334801, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.045222170650959015, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.028650468215346336, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.024489758536219597, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.024316884577274323, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.022499648854136467, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.021959513425827026, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.016015779227018356, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.016248371452093124, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.015067745000123978, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.011971552856266499, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.11140166968107224, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.11140166968107224, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.03743519261479378, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.03332744166254997, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.029433421790599823, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.025646470487117767, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.01753845438361168, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.014724504202604294, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.02449243888258934, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.022040219977498055, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.017925016582012177, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.01460693683475256, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.014252508990466595, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.012493953108787537, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.010629557073116302, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.008613953366875648, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.008075535297393799, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.006295450963079929, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.004681159742176533, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.004337879829108715, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.004113632254302502, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.00377477565780282, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0033908758778125048, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.003431598888710141, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0027804067358374596, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.002463374985381961, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.03743519261479378, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.03743519261479378, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.038030970841646194, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.03339137136936188, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.028564145788550377, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.025009218603372574, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.01740424335002899, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.013992778025567532, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.02550419420003891, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.02298979088664055, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.017842082306742668, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.01454109512269497, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.01423387136310339, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.012863130308687687, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.010976502671837807, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.008482053875923157, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.007796244230121374, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.006408471614122391, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.004544364754110575, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.004089607857167721, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.003997479565441608, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.0035442740190774202, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.003369581187143922, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.0033643688075244427, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.00251317722722888, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.002262231893837452, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.038030970841646194, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.038030970841646194, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.19701915979385376, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.1779184192419052, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.16999468207359314, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.14951595664024353, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.0924171656370163, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.08462385088205338, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.10952644795179367, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.09931164234876633, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.09393927454948425, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.07828203588724136, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.07391881942749023, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.05610549449920654, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.047477833926677704, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.044307515025138855, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.04352784901857376, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.02809375897049904, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.022456007078289986, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.021790649741888046, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.01950383372604847, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.01897556520998478, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.014282786287367344, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.013409015722572803, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.012821614742279053, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.007960902526974678, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.10952644795179367, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.10952644795179367, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.18265022337436676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.14905177056789398, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.13624942302703857, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11335214972496033, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08403900265693665, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07097455859184265, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10145974159240723, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09217353165149689, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08655057847499847, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.062481172382831573, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.057467833161354065, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05232985317707062, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04498709365725517, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04135200381278992, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04045247659087181, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.026722749695181847, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.022979188710451126, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02213146537542343, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.019170379266142845, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01859946921467781, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015385997481644154, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016482295468449593, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014286014251410961, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013165788725018501, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11335214972496033, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11335214972496033, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.22557947039604187, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21220210194587708, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.20785699784755707, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18814536929130554, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10608632862567902, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10158128291368484, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11784013360738754, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10868675261735916, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10697408765554428, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0950264111161232, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09058107435703278, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.060012731701135635, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05206012353301048, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05083366483449936, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05055736377835274, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030069904401898384, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02631736919283867, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026016656309366226, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02410886436700821, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.023932715877890587, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016039112582802773, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016110733151435852, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01564013585448265, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011004811152815819, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11784013360738754, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11784013360738754, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.25898507237434387, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.24388381838798523, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2389926016330719, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.21641471982002258, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12174438685178757, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11665821075439453, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.13491374254226685, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1247396469116211, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12278290838003159, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.1091218963265419, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10399686545133591, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06861498951911926, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05953984335064888, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05817513167858124, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.0578562431037426, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03427678346633911, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.029601365327835083, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.029246464371681213, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.026997525244951248, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.026795020326972008, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01784905232489109, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01733546517789364, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.017379460856318474, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010848021134734154, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.1091218963265419, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.1091218963265419, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.11438062787055969, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.10213244706392288, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.09327191114425659, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.08313842862844467, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.05189353972673416, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.04599585756659508, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.06900178641080856, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.06266926229000092, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.05370228365063667, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.045092277228832245, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.04254325479269028, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.033170733600854874, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.029180515557527542, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.02531423047184944, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.024216922000050545, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.016863690689206123, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.013566341251134872, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.013108323328197002, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.012279577553272247, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.011523048393428326, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.009294652380049229, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.009500990621745586, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.00804077833890915, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.0067343441769480705, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.11438062787055969, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.11438062787055969, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.06332339346408844, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.05757429823279381, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.054081693291664124, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.04787634313106537, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.02973179891705513, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.026809638366103172, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.03624027594923973, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.03346862271428108, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.030217941850423813, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.025498399510979652, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.024234643206000328, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.018521009013056755, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.01608394645154476, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01432099286466837, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.013882969506084919, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.009275957942008972, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.0074408468790352345, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.00712543586269021, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.006597401574254036, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.00630901288241148, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.004850080236792564, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0048398650251328945, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0042513152584433556, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.003182054031640291, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.06332339346408844, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.06332339346408844, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07192264497280121, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.0643751472234726, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.06012450158596039, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.052720773965120316, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.03327716514468193, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.02964801713824272, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.04094409570097923, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.03766077756881714, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.03389640152454376, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.02802315168082714, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.026632104068994522, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.020706215873360634, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.01798289641737938, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.015993250533938408, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.015477904118597507, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.010353153571486473, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.008175498805940151, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.007804662920534611, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.007104469928890467, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.006764892488718033, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.005330887157469988, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.005171317607164383, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.004635274410247803, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0031641165260225534, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07192264497280121, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07192264497280121, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.2107476443052292, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.19322524964809418, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.18625646829605103, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.16545581817626953, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.09902696311473846, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.09213828295469284, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11407171934843063, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.10493841767311096, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.10037371516227722, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.08556874096393585, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.08085848391056061, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.05813949182629585, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.050103772431612015, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.04740789160132408, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.0467698909342289, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.029101606458425522, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.02406802587211132, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.023491863161325455, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.021258430555462837, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.020823035389184952, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.014874021522700787, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.014286678284406662, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.013789999298751354, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.00860561616718769, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11407171934843063, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11407171934843063, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.20506733655929565, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.16987957060337067, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.15625737607479095, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12460354715585709, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09544305503368378, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08199670910835266, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11711984127759933, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10387865453958511, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09779707342386246, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06906570494174957, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06487732380628586, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.059891168028116226, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05110607668757439, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.047360192984342575, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04644651338458061, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03133465349674225, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.026912696659564972, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02606077864766121, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.022415030747652054, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02181898057460785, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.019014116376638412, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019807133823633194, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017911940813064575, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01631208509206772, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11711984127759933, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11711984127759933, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2417636662721634, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22717058658599854, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2225237339735031, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20119455456733704, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11406899243593216, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10910637676715851, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1264764815568924, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11687100678682327, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11507241427898407, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.1018129512667656, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09691710770130157, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06433680653572083, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0559135340154171, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05464942753314972, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05436098203063011, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03221730887889862, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.028170177713036537, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.027848126366734505, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025697246193885803, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02550547383725643, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016986219212412834, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.017044102773070335, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016569038853049278, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01140655018389225, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11687100678682327, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11687100678682327, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2785564064979553, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2617899179458618, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2565643787384033, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.2320770025253296, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.13142140209674835, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.12574811279773712, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.14542178809642792, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.13454477488994598, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.13255302608013153, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11738330870866776, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.11160755157470703, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.07379491627216339, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.06421472132205963, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.06279382109642029, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.0624561607837677, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03685763478279114, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.0318794772028923, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.03151172772049904, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.028976989910006523, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02876192145049572, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.019032446667551994, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.018564732745289803, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.018559126183390617, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011458156630396843, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11738330870866776, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11738330870866776, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.25743478536605835, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.23153939843177795, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.22174009680747986, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19994297623634338, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11753653734922409, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10855545103549957, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13788293302059174, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.126409113407135, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12131012976169586, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10337325185537338, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09756104648113251, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06989564746618271, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06062757968902588, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.056633368134498596, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05567900836467743, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03510482609272003, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.029850903898477554, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.02966287173330784, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.027105208486318588, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.026387399062514305, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01895272731781006, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01943461410701275, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017651639878749847, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013841300271451473, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12131012976169586, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11753653734922409, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.06443937122821808, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.05931231752038002, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.05565643683075905, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.04955591633915901, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.030382506549358368, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.027509236708283424, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.03736312687397003, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.03461504355072975, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.030810141935944557, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.026446998119354248, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.025202816352248192, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.01906881108880043, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.01665203832089901, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01471068523824215, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.014221693389117718, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.00956894550472498, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.007779226638376713, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.007445854134857655, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.007015001028776169, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.006707187741994858, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.005076719913631678, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.00524775218218565, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.004459287505596876, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.003668236080557108, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.06443937122821808, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.06443937122821808, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.06042911112308502, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.05523372441530228, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.051389697939157486, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.04554392397403717, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.028151823207736015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.02524111047387123, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.03509887680411339, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.03249175101518631, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.028567293658852577, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.02429600991308689, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.02313421294093132, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.017677707597613335, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.015519307926297188, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.013517037965357304, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.013022532686591148, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.00882752239704132, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.0069656712003052235, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.006624938920140266, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.006190737243741751, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.005874703638255596, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.004591500386595726, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.00452544167637825, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.003976072650402784, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0028566191904246807, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.06042911112308502, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.06042911112308502, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.23218286037445068, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.21403875946998596, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.20734557509422302, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.18445217609405518, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.1090502217411995, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.10214686393737793, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.12575092911720276, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11411906033754349, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11053379625082016, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.09476129710674286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.08976669609546661, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06413249671459198, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.05457291379570961, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05228692293167114, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.0517120435833931, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.032108936458826065, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.026506522670388222, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.025964470580220222, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.0235306229442358, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.02314906381070614, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.016424594447016716, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.015618021599948406, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01527592446655035, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0094336848706007, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11411906033754349, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11411906033754349, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.17685294151306152, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.15406355261802673, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.14515599608421326, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11647310853004456, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08143898844718933, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07289251685142517, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09707196801900864, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.08886751532554626, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08410327881574631, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06412162631750107, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.057320527732372284, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05013028904795647, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.043344464153051376, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.039947833865880966, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0391353964805603, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.025468163192272186, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02206920087337494, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.021411878988146782, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.018971219658851624, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.018442070111632347, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01443847268819809, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.015681592747569084, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.013402917422354221, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.012405660934746265, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11647310853004456, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11647310853004456, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.22359852492809296, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.20990276336669922, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.205328106880188, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18537352979183197, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10594696551561356, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.1011749729514122, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11758335679769516, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10883898288011551, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10695996880531311, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09437281638383865, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0897219181060791, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06008218228816986, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.052274368703365326, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.050983164459466934, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05068698525428772, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03005100041627884, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026628000661730766, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026306618005037308, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02433042787015438, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02414592169225216, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016023024916648865, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01665477827191353, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015598531812429428, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011770948767662048, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11758335679769516, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11758335679769516, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.27747994661331177, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2605046033859253, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.25494781136512756, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.23034484684467316, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.13127541542053223, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.1253950446844101, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.145718052983284, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.13472050428390503, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.13244298100471497, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11693134158849716, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.11116431653499603, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.07403059303760529, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.06435336172580719, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.06278662383556366, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.062417663633823395, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.037004344165325165, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.03194611892104149, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.031538255512714386, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.028975967317819595, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02873942442238331, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.019202198833227158, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.018709518015384674, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.0186676774173975, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011658422648906708, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11693134158849716, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11693134158849716, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24665915966033936, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.21857735514640808, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20837220549583435, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18588034808635712, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.1123887225985527, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.1027936339378357, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13128790259361267, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12111025303602219, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11646447330713272, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.0969693511724472, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09094611555337906, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06702423840761185, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05821925029158592, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05427028611302376, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.0533120259642601, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.033770643174648285, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.028755372390151024, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028604310005903244, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.025791063904762268, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025072554126381874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018422747030854225, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.018933573737740517, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01719764620065689, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013695038855075836, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12111025303602219, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12111025303602219, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07134174555540085, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.06554526835680008, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.06201116740703583, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.05518077313899994, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03363038972020149, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.030688541010022163, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04054129123687744, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.037431471049785614, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0341060608625412, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.02913525141775608, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0276933666318655, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.02067527174949646, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.017939487472176552, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.016208987683057785, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.015781093388795853, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.010340792126953602, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.008455454371869564, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.008143141865730286, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.007560040336102247, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.007283533923327923, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.005421033129096031, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.005486424081027508, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.004831667058169842, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.003690549172461033, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07134174555540085, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07134174555540085, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07057995349168777, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.06403092294931412, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.06024453043937683, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.053148381412029266, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.03273513540625572, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.02961108274757862, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.03993598371744156, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.03685860335826874, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.033304497599601746, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.02807190828025341, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.026673289015889168, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.020258110016584396, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.017603006213903427, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.015777522698044777, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.015301109291613102, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.010153635405004025, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.008135680109262466, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.007799661718308926, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.007178628817200661, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.006888618227094412, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.005254840012639761, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.005199046805500984, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.004618953913450241, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0033402983099222183, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07057995349168777, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07057995349168777, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.21940584480762482, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.20250903069972992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.19610869884490967, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.1748703122138977, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.10357871651649475, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.09691504389047623, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11910523474216461, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.10887662321329117, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.10484466701745987, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.09014247357845306, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.085489921271801, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06071928143501282, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.052185915410518646, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.049642689526081085, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.04902143403887749, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.030412010848522186, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.025123264640569687, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.024574339389801025, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.022332295775413513, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.021920757368206978, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.015545934438705444, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.014817317947745323, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01446698047220707, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.00885627418756485, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11910523474216461, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11910523474216461, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.18718932569026947, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1621657758951187, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.15221722424030304, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12489593029022217, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08686425536870956, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07686102390289307, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10467781126499176, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09422460943460464, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08859711140394211, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06835513561964035, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06208515167236328, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.053731974214315414, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04601113870739937, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04268631711602211, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.041877057403326035, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.027526305988430977, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02359093725681305, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.022905655205249786, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02028835006058216, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01977137289941311, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015952669084072113, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01674991101026535, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014949331991374493, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013294353149831295, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10467781126499176, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10467781126499176, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.20963609218597412, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1967342644929886, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1923506259918213, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.17379948496818542, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09933458268642426, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09476719051599503, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11052794754505157, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10223987698554993, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10023906081914902, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08847659826278687, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08412587642669678, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05644078925251961, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.049102455377578735, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04780762270092964, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04749847576022148, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02824004739522934, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02499481849372387, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.024682674556970596, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.022843532264232635, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.022657962515950203, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.0150977224111557, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015685899183154106, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01467160414904356, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011127840727567673, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11052794754505157, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11052794754505157, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.27025166153907776, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2537742555141449, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2483244091272354, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.22444862127304077, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1278429478406906, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.12202892452478409, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.14206314086914062, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.13134929537773132, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12902472913265228, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11392933130264282, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10831138491630554, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.07223980128765106, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.06276408582925797, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0611676350235939, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.060790978372097015, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03605029359459877, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.031100254505872726, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.03070310316979885, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.028216926380991936, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.027974147349596024, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.018650971353054047, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.018220262601971626, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.018102269619703293, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011334896087646484, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11392933130264282, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11392933130264282, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24133236706256866, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2151346653699875, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.205501526594162, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18309111893177032, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11022786051034927, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10131826996803284, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.12927600741386414, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11840389668941498, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11387445032596588, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09553661942481995, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08947168290615082, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06585682928562164, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05707007274031639, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05339157581329346, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.052489690482616425, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.033291637897491455, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.028614990413188934, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028463559225201607, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02583838440477848, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025177378207445145, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018493453040719032, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019218923524022102, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017320945858955383, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014377596788108349, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11840389668941498, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11840389668941498, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07926647365093231, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.07379167526960373, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.07071942090988159, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.06348370015621185, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03744921460747719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.0348266139626503, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04385857656598091, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.04070327430963516, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.03787597268819809, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.03305675461888313, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.031477414071559906, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.022388463839888573, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.019543621689081192, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01799960248172283, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0176189374178648, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.011212339624762535, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.009270722977817059, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.00899257231503725, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.00837669800966978, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.008131580427289009, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.005834618117660284, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0058046686463057995, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.005324193276464939, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0037406934425234795, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07926647365093231, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07926647365093231, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.076785147190094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.07111968100070953, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.06794434785842896, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.060621477663517, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.03585813567042351, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.03324891999363899, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.04238280653953552, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.03917449712753296, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.03630046546459198, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.03140418976545334, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.029830386862158775, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.02135256677865982, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.01869957521557808, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.017236867919564247, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.016881410032510757, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.010708657093346119, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.0088302968069911, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.008554150350391865, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.007920202799141407, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.007685557473450899, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.005559628829360008, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.005477238912135363, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.005087230820208788, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0034517187159508467, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.076785147190094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.076785147190094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.23749670386314392, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.22042638063430786, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.2141454666852951, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.19172920286655426, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.11167817562818527, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.10529834777116776, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.12706232070922852, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11641138792037964, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11295794695615768, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.09794361889362335, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.09313234686851501, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.0648428201675415, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.05563421547412872, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.053421590477228165, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.052867159247398376, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03243054077029228, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.026997480541467667, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.026497242972254753, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.024136608466506004, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.02377156913280487, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.01652396097779274, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.01572355069220066, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.015518603846430779, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.009282099083065987, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11641138792037964, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11641138792037964, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.21411347389221191, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.19089831411838531, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.18272094428539276, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.15531966090202332, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0991462916135788, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09083860367536545, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11446603387594223, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10496459156274796, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10117066651582718, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08157722651958466, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0751064345240593, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05875233933329582, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05085485056042671, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.048195723444223404, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04756518453359604, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02962224930524826, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02598581649363041, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.025411415845155716, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.022687865421175957, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.022288085892796516, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.016456175595521927, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017550911754369736, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015663091093301773, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013409440405666828, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11446603387594223, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11446603387594223, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.20035649836063385, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18806317448616028, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18395376205444336, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16632142663002014, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09473411738872528, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09041944891214371, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1053563579916954, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09739410132169724, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09560047090053558, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08443288505077362, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08035038411617279, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05372563377022743, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04670846462249756, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04550967365503311, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04523099213838577, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.026881089434027672, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.023652393370866776, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.023365342989563942, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02160809189081192, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02142680250108242, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014297645539045334, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01465601846575737, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013894125819206238, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01017626840621233, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1053563579916954, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1053563579916954, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2701250910758972, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.253938227891922, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.24854740500450134, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.22480590641498566, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12765350937843323, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.12197276204824448, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1418173462152481, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.13111943006515503, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12884971499443054, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11395937204360962, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10840793699026108, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.07210028171539307, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.06262440979480743, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.06106511875987053, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.060695063322782516, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03600887209177017, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.031077805906534195, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.030686594545841217, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.028247786685824394, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.028010616078972816, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.0186769999563694, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.018238522112369537, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01813350059092045, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011393509805202484, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11395937204360962, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11395937204360962, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.23685303330421448, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.21172292530536652, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20243901014328003, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18091510236263275, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.10849804431200027, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.09978672116994858, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.12715253233909607, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11664239317178726, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11213532835245132, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09441763162612915, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08858513087034225, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.0649646446108818, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05627286434173584, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05258976295590401, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.0517168827354908, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.032861147075891495, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.028182458132505417, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028003325685858727, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02551271952688694, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.0248318649828434, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01829337887465954, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.018915176391601562, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017119668424129486, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014081341214478016, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11664239317178726, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11664239317178726, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07596584409475327, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.07007786631584167, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.06714461743831635, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.06019497662782669, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03577350080013275, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.033170185983181, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04317223280668259, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.038829416036605835, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0365462601184845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.031398218125104904, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03005465865135193, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.02210678905248642, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.018711023032665253, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.017226319760084152, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.016858788207173347, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0110857542604208, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.008943883702158928, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.00870716292411089, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.008065497502684593, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.007833361625671387, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0058261011727154255, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.005686208605766296, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0052068340592086315, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0037927585653960705, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07596584409475327, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07596584409475327, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07450142502784729, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.06856897473335266, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.0650126039981842, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.05793570354580879, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.03471757471561432, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.03185534104704857, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.041378818452358246, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.038539692759513855, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.03517580032348633, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.030225226655602455, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.02868049219250679, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.020919224247336388, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.018353354185819626, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.016665730625391006, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.01624971441924572, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.010470039211213589, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.008503856137394905, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.008197026327252388, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.007573944982141256, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.007300129160284996, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.005405707284808159, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.005258585792034864, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.004869810305535793, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.003203015774488449, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07450142502784729, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07450142502784729, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.2353893667459488, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.21803821623325348, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.2112506926059723, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.18888431787490845, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.11101846396923065, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.10416755825281143, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.12755422294139862, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11657769232988358, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11228226125240326, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.0970175489783287, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.0922975018620491, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06530538201332092, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.055780135095119476, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05311279371380806, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05246582627296448, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.032614585012197495, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.026892174035310745, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.026330573484301567, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.023996837437152863, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.02356436289846897, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.016703322529792786, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.01582438498735428, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.015517564490437508, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.009425155818462372, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11657769232988358, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11657769232988358, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.21295587718486786, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1889917552471161, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.17913900315761566, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1551329493522644, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09890538454055786, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08928856253623962, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11687145382165909, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10697543621063232, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10106539726257324, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08197372406721115, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07700005918741226, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06022997945547104, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05221540480852127, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04857683554291725, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04770473763346672, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.030617602169513702, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02687690407037735, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02616608515381813, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.023813728243112564, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02327900007367134, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017501138150691986, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019074922427535057, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016451748088002205, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.015187138691544533, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11687145382165909, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11687145382165909, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.22541086375713348, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21159790456295013, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2069171965122223, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18714767694473267, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1073119044303894, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10241026431322098, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11981233954429626, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11054578423500061, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10830636322498322, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09576378762722015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09133300185203552, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06175047904253006, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05386556312441826, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05239688977599144, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.052061788737773895, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.0311118196696043, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.028600161895155907, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.028272034600377083, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02646738477051258, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.026254970580339432, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01748507469892502, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.019516926258802414, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.017008356750011444, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.015449853613972664, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11981233954429626, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11981233954429626, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2796771824359894, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2628765106201172, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.25726884603500366, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.23283974826335907, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.13245642185211182, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.1264609694480896, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.14721347391605377, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.13615109026432037, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.13366086781024933, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11815381050109863, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.11253546923398972, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.07503169775009155, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.06514637172222137, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.06339924782514572, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.06300274282693863, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.037528056651353836, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.032436441630125046, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.03201577067375183, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.029507046565413475, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02924523688852787, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01963072270154953, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.019310522824525833, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.019043125212192535, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.012404488399624825, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11815381050109863, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11815381050109863, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.09456131607294083, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.0938773974776268, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.029855456203222275, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.029491711407899857, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.0285917017608881, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.005406854674220085, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.09626249969005585, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.09305036067962646, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.028744246810674667, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.028433475643396378, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.028503689914941788, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.028494346886873245, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.02819906361401081, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.007269924972206354, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.003377942368388176, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.006890760734677315, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.002534301718696952, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.0017085266299545765, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.0024377244990319014, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.0015402451390400529, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.002362702740356326, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.002286989241838455, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.0011848267167806625, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.0011644762707874179, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.09456131607294083, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.09456131607294083, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.08846572041511536, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.08207321912050247, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.07893195748329163, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.07071974128484726, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.041867442429065704, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.03904131054878235, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.048856042325496674, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.04491487145423889, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04236576333642006, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0367201529443264, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.034928709268569946, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.024861525744199753, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.021527517586946487, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.020129874348640442, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.019792115315794945, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.012426195666193962, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.010401111096143723, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.010138357058167458, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.009354151785373688, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.009138708934187889, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0065009696409106255, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.006488627754151821, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006022689864039421, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004281836096197367, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.08846572041511536, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.08846572041511536, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08551570028066635, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.07903925329446793, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.07510839402675629, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.06694434583187103, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.03995528817176819, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.036886006593704224, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.0476461797952652, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04410813748836517, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.0405128188431263, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.03485849127173424, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.033085450530052185, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.02398393489420414, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.020943840965628624, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.019183073192834854, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.01873566024005413, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.012001191265881062, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.009822099469602108, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.009483478963375092, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.008772923611104488, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.008475340902805328, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006222590338438749, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006085934583097696, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.0056480709463357925, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.00376211223192513, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08551570028066635, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08551570028066635, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.25713443756103516, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.23912093043327332, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.23254172503948212, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.2082996368408203, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.1214953064918518, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.1147637665271759, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.1380244493484497, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1262437105178833, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.12296222895383835, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10653718560934067, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.10128489881753922, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.0704524964094162, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.060303233563899994, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05816090852022171, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05764499306678772, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03519962728023529, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.02933734469115734, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.028832970187067986, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.026186082512140274, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.02583996020257473, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.01793098635971546, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.016944434493780136, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.016901668161153793, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.009930131956934929, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.1214953064918518, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.1214953064918518, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.21493150293827057, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1912865936756134, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.18251915276050568, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.16417790949344635, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09995709359645844, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09053304046392441, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11539969593286514, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10618242621421814, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10196976363658905, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08437322825193405, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08013299107551575, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.059251319617033005, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.0512334443628788, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04842051863670349, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.047745056450366974, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.029701190069317818, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.025757556781172752, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.025148434564471245, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.022763263434171677, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02234686352312565, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.016170473769307137, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01702331379055977, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0152695057913661, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.012599538080394268, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11539969593286514, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11539969593286514, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.21192054450511932, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.19882263243198395, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.19443605840206146, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.17553074657917023, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.100309357047081, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09568867832422256, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11182492971420288, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10326126962900162, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10127446800470352, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08927156031131744, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08494701236486435, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.056938886642456055, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04943114519119263, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0481138601899147, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.047806158661842346, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.028557991608977318, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.024781232699751854, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.024466130882501602, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02253604866564274, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.022343583405017853, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015102650038897991, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015005598776042461, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.014664726331830025, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009989437647163868, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11182492971420288, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11182492971420288, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2787718176841736, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2617291808128357, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.25612303614616394, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.23147201538085938, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.13208825886249542, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.12606796622276306, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1465795636177063, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1356946974992752, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.13332407176494598, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11763148009777069, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.11181598901748657, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.07454149425029755, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0648297592997551, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0631871223449707, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.06279950588941574, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03721928223967552, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.03211977332830429, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.03170240297913551, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.029123026877641678, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.028872765600681305, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.019221434369683266, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.018791869282722473, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.018659014254808426, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01165022049099207, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11763148009777069, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11763148009777069, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2544926404953003, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2278284877538681, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21778792142868042, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19475027918815613, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11696688830852509, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10750627517700195, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13724906742572784, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1258525848388672, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12082865089178085, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10161034762859344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09524492174386978, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06982320547103882, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06045155972242355, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.056429505348205566, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.055446989834308624, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03518885001540184, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.029747193679213524, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029568158090114594, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026765594258904457, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.026020018383860588, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01919594407081604, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019381916150450706, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01790488325059414, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013788571581244469, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12082865089178085, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12082865089178085, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.0697208046913147, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.0643630176782608, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.061524853110313416, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.0549355112016201, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03297223150730133, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.03049791231751442, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.03913760930299759, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.03581302613019943, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.033414095640182495, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.028749514371156693, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.02737008035182953, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.019924718886613846, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.01725856401026249, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01594294235110283, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.015623848885297775, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01000871043652296, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.008383591659367085, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.008143982850015163, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.007550956215709448, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.007349771447479725, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.005307919345796108, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.005463168025016785, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.004859934560954571, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0038412907160818577, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.0697208046913147, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.0697208046913147, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07810463756322861, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.07166469842195511, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.06780651211738586, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.06021959334611893, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.036397580057382584, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.03320645913481712, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.044046152383089066, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04064495489001274, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.036892663687467575, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.03153280168771744, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.030015243217349052, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.022268373519182205, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.019403638318181038, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.01750222221016884, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.017004042863845825, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.011144272983074188, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.009004347957670689, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.00865353923290968, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.008023814298212528, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.00770579231902957, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.005783041473478079, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.005711445119231939, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.005165064241737127, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0036346616689115763, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07810463756322861, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.07810463756322861, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.2201356142759323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.20374202728271484, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.1973833292722702, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.17641347646713257, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.10365477949380875, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.09736630320549011, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11968903243541718, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.10892514139413834, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.10499674081802368, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.09063726663589478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.08609654009342194, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.061059702187776566, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.05207029730081558, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.04967067763209343, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.04907971993088722, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.030489254742860794, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.025126969441771507, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.024601725861430168, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.022421834990382195, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.022037744522094727, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.015570015646517277, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.014797358773648739, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.014446264132857323, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.008814520202577114, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11968903243541718, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11968903243541718, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.20762142539024353, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.17981718480587006, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.16534313559532166, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.14177583158016205, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09566706418991089, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08326825499534607, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1237446591258049, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1101253554224968, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0987166315317154, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07800479233264923, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07415292412042618, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06365750730037689, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05379186198115349, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.047228217124938965, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0455700159072876, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.032691240310668945, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.026534095406532288, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.025450287386775017, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02342168241739273, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.022420700639486313, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01901099644601345, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01963140442967415, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01694844290614128, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.015628209337592125, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1101253554224968, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1101253554224968, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.20458996295928955, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.19206541776657104, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18787722289562225, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16974839568138123, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0970141664147377, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09258583933115005, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10782218724489212, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09981836378574371, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.0979064404964447, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08646132797002792, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0821835920214653, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05496616289019585, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04783723130822182, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04659218341112137, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.046291984617710114, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02747611328959465, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.024116331711411476, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02381405234336853, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.021992934867739677, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.021799158304929733, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014492920599877834, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014780641533434391, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.014079932123422623, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01007247157394886, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10782218724489212, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10782218724489212, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2756986618041992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2592776119709015, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.25377586483955383, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.22965693473815918, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.13077232241630554, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.12491913139820099, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.14526066184043884, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.13419482111930847, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.1319330483675003, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11665146797895432, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.11100045591592789, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.07383140176534653, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0641409307718277, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.06257142126560211, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.06220421940088272, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.036897171288728714, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.03184143826365471, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.03144237399101257, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.028926439583301544, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.028684964403510094, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01916315406560898, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.018677968531847, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.018623843789100647, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011656136251986027, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11665146797895432, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11665146797895432, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2642398476600647, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.235836923122406, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.22582094371318817, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.20113223791122437, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12155022472143173, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.11171294003725052, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14138293266296387, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12995167076587677, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12546484172344208, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10492117702960968, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09809944778680801, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07207279652357101, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06246134266257286, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.058663297444581985, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05775153636932373, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.036247219890356064, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.030962670221924782, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.030825067311525345, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.027737755328416824, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02703593298792839, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019657272845506668, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.02016017958521843, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.018438851460814476, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014454521238803864, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10492117702960968, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10492117702960968, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.09196139127016068, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.08483471721410751, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.08125782757997513, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.07246057689189911, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0432903878390789, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.040163375437259674, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05081793665885925, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.04675538092851639, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04384171962738037, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.03772937133908272, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03578509762883186, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.025887465104460716, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.022455498576164246, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0208817757666111, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02049916237592697, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.012959077022969723, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.010908353142440319, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.010609212331473827, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.009802162647247314, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.009551293216645718, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.006823252886533737, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007000455167144537, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006272564642131329, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004814774263650179, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.09196139127016068, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.09196139127016068, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09054621309041977, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08283481746912003, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.07898234575986862, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.06997141987085342, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.042035169899463654, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.0385899618268013, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.04979797080159187, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.045797280967235565, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04263816773891449, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.0363340862095356, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03446199372410774, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.025221342220902443, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.021953653544187546, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.020185371860861778, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.019762404263019562, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.012646576389670372, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.010440521873533726, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.010099527426064014, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009310821071267128, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.009015779942274094, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.0066034626215696335, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006598328240215778, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.0059990533627569675, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004322068300098181, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09054621309041977, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09054621309041977, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.22797167301177979, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.21211683750152588, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.2062118947505951, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.18471169471740723, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.10769592225551605, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.10175768285989761, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.12254992127418518, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11214186251163483, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.10884958505630493, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.09460736066102982, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.08978129178285599, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06256280839443207, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.053631022572517395, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05159154534339905, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05109022557735443, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.031224438920617104, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.026125213131308556, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.025659743696451187, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.023401619866490364, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.023073142394423485, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.016037169843912125, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.01530465018004179, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.015135006979107857, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.009218418970704079, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11214186251163483, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11214186251163483, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22916726768016815, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20596501231193542, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.19620218873023987, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.16998130083084106, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10678618401288986, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09739330410957336, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12586244940757751, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11480899900197983, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10872235894203186, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08947630226612091, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.083041250705719, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0648249089717865, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05559039115905762, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.051960039883852005, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05109516531229019, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.032785046845674515, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02799253724515438, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027265038341283798, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.024689795449376106, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.024136211723089218, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01836671307682991, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01897444948554039, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01726529933512211, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014358403161168098, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11480899900197983, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11480899900197983, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1862482726573944, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17478446662425995, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.17073804140090942, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15429693460464478, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08834460377693176, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08419017493724823, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09852585941553116, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09116985648870468, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.0891808271408081, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0786515548825264, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07475853711366653, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0501951202750206, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.043697141110897064, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.042433224618434906, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04212813451886177, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.025116894394159317, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021977044641971588, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.021686557680368423, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02002859115600586, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.019838180392980576, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013253431767225266, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013512286357581615, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012836175039410591, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009211347438395023, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09852585941553116, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09852585941553116, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.274025022983551, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.25745540857315063, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2518984079360962, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.22779083251953125, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12997837364673615, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.1240624189376831, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.14457450807094574, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.133628711104393, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.13114754855632782, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11586897820234299, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.11014764755964279, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.07353491336107254, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0638885423541069, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.06222224608063698, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.061821237206459045, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.036716386675834656, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.031684666872024536, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.03128274530172348, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.028763053938746452, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.028509773313999176, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01900409534573555, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.018633868545293808, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.018427370116114616, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011677726171910763, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11586897820234299, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.11586897820234299, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.26599323749542236, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.23826918005943298, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.22820788621902466, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.20267891883850098, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12258835136890411, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.11295828223228455, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14255261421203613, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.13111715018749237, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.1264675408601761, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10590220242738724, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09880062192678452, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07274043560028076, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06299731135368347, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.0591449961066246, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05820335075259209, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.036562204360961914, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.031173599883913994, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.031035421416163445, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.027934202924370766, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.027219083160161972, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019813193008303642, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.020266655832529068, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.0185826625674963, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014487138018012047, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10590220242738724, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10590220242738724, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07894766330718994, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.07349061220884323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.07078863680362701, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.06352085620164871, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03731444850564003, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.03493332490324974, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.043402571231126785, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.03999210149049759, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.03774292767047882, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.032874926924705505, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03124522976577282, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.02210780419409275, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.019137822091579437, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01796182058751583, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.017671342939138412, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.011049924418330193, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.009308713488280773, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.00908865500241518, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.00841487292200327, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.008231871761381626, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.005770568270236254, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0058469390496611595, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.005347536876797676, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.003924879711121321, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07894766330718994, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07894766330718994, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.06948234885931015, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.06445442140102386, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.06114248186349869, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.054792147129774094, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.032403819262981415, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.029856903478503227, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.03869696706533432, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.035986870527267456, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.03278088569641113, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.028489384800195694, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.027026940137147903, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.019410161301493645, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.017093656584620476, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.015502624213695526, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.015131118707358837, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.00970324594527483, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.007929876446723938, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.007657285779714584, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.007128030527383089, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.006878441199660301, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.005023171193897724, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.0049231993034482, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.004541890695691109, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0030333432368934155, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.06948234885931015, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.06948234885931015, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.24863605201244354, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.23175086081027985, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.22588326036930084, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.2021268755197525, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.11738016456365585, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.11123127490282059, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.13279879093170166, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1216333657503128, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11865280568599701, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10316746681928635, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.0979415625333786, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06776086986064911, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.05809512361884117, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.056126002222299576, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.055646974593400955, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.033899396657943726, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.02837318554520607, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.02790827862918377, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.025412507355213165, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.025116274133324623, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.017308395355939865, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.01643224060535431, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.016355644911527634, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.009767529554665089, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11865280568599701, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11865280568599701, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.21687398850917816, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.19433587789535522, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.18401020765304565, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1571563482284546, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09991037100553513, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09062882512807846, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12162579596042633, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11037780344486237, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10254260152578354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08405306935310364, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07709480077028275, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06283807754516602, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.053618982434272766, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.048890963196754456, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04765842854976654, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03211861848831177, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.026700537651777267, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02588568441569805, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.023774603381752968, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.023048853501677513, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01815417781472206, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018754340708255768, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016521302983164787, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014440568163990974, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11037780344486237, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11037780344486237, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1711510717868805, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.16070036590099335, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.15696170926094055, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14198385179042816, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08112423866987228, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07732988148927689, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09061159938573837, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08377335965633392, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08187313377857208, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07231131941080093, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.06889647990465164, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04620542377233505, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04025166481733322, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03906567394733429, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.038792673498392105, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02314808778464794, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.020447196438908577, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.020173585042357445, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01870705373585224, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.018530208617448807, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01237491238862276, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012874800711870193, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.011994161643087864, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.0091338399797678, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09061159938573837, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09061159938573837, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2541617751121521, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.23893558979034424, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.23384541273117065, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.2115514874458313, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12047223746776581, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11504167318344116, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1341908723115921, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12393596023321152, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12155698239803314, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10753291100263596, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10235131531953812, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06830966472625732, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05925330892205238, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05768068879842758, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05731938034296036, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.034114208072423935, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02942080795764923, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.029036816209554672, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02674918994307518, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02651062048971653, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.017718536779284477, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.017366399988532066, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.017162976786494255, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01096719317138195, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12047223746776581, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12047223746776581, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.26346874237060547, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2371780127286911, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.22806890308856964, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.20282220840454102, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12183044105768204, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.11291110515594482, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1407666653394699, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12927021086215973, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12535414099693298, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10577595978975296, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.0985676646232605, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07209805399179459, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.062291644513607025, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05893000587821007, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05811689421534538, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03625097870826721, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.031313952058553696, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.0312051922082901, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02823401242494583, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.027600880712270737, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01978680118918419, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.02062164805829525, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.018682757392525673, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.015164356678724289, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10577595978975296, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10577595978975296, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.09939330071210861, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09208990633487701, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.0887337252497673, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.07939580082893372, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.046767741441726685, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.043697040528059006, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.054105374962091446, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.049913547933101654, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04731693118810654, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.040959253907203674, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0388757660984993, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.027586378157138824, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.023939261212944984, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.022502725943922997, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02214519865810871, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01381063461303711, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01169116236269474, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011413633823394775, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010533162392675877, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010311762802302837, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007249828893691301, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007374992594122887, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006745188031345606, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005011390428990126, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.09939330071210861, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.09939330071210861, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09004811942577362, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08290836215019226, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.07941966503858566, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07067796587944031, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04179103672504425, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.03867057338356972, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.04888070002198219, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.045154765248298645, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04229690879583359, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.036444731056690216, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.034478940069675446, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.0247826986014843, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.021603593602776527, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.020018963143229485, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.019646627828478813, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.012397143058478832, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.010280718095600605, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.00998126994818449, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009194068610668182, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.00893991719931364, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006429844070225954, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006365568842738867, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.005906044505536556, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004054481629282236, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09004811942577362, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09004811942577362, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.22322416305541992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.2080802470445633, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.20266523957252502, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.18175165355205536, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.10514387488365173, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.09960811585187912, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11939623951911926, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.10916902124881744, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.10632230341434479, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.09275476634502411, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.08800041675567627, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.0607159323990345, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.052155688405036926, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05025193467736244, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.0498613566160202, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03029809333384037, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.025431951507925987, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.024998141452670097, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.022815564647316933, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.02252287045121193, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.015488282777369022, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.014790132641792297, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.014697945676743984, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.008774522691965103, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11939623951911926, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.11939623951911926, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.23443444073200226, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20763137936592102, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1982414573431015, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1663978099822998, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.1085413321852684, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09883461147546768, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12605148553848267, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11482590436935425, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11039318144321442, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08803682774305344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08019870519638062, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0645727664232254, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05553535744547844, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05270242318511009, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05203396826982498, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0324983112514019, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02827857993543148, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027643706649541855, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02439969964325428, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.023977117612957954, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017730575054883957, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018923450261354446, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016841359436511993, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014300048351287842, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11482590436935425, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11482590436935425, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1765177845954895, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.16574037075042725, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16189172863960266, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14644622802734375, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08365815132856369, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07974671572446823, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09339964389801025, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08630869537591934, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.0844140499830246, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07454940676689148, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07100676745176315, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04761649668216705, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04142698273062706, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0402328222990036, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.039953459054231644, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02385193109512329, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02097809687256813, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.020702257752418518, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019167162477970123, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.018987376242876053, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012729745358228683, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.0130923418328166, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01233357097953558, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009160392917692661, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09339964389801025, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09339964389801025, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2540082633495331, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.23881950974464417, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.23358604311943054, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.21140335500240326, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12041983008384705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11499637365341187, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1341993808746338, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1239045262336731, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12153736501932144, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.1074451357126236, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10226838290691376, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06820929050445557, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05921191722154617, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05765580013394356, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05728644132614136, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.034084875136613846, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.0294208787381649, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.029033251106739044, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02673708088696003, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.026493780314922333, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.017733585089445114, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.017386576160788536, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01718878746032715, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011017200537025928, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12041983008384705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12041983008384705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.267843097448349, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2414059191942215, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.23202137649059296, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.20646987855434418, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12408735603094101, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.11484630405902863, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14338424801826477, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.131749227643013, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12766215205192566, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10766557604074478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.10046499222517014, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07340353727340698, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06350080668926239, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.06003408879041672, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.059200093150138855, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.036879632622003555, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.0318894125521183, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.03176981583237648, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02873433753848076, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.028090493753552437, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.020053278654813766, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.020987972617149353, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.018958333879709244, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.015417939983308315, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10766557604074478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10766557604074478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10354389250278473, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09606947004795074, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09231901168823242, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08260076493024826, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04885590076446533, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04557352513074875, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.057023074477910995, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05241796001791954, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.049408938735723495, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04281458258628845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04073188826441765, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.029077591374516487, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.025150347501039505, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.023506266996264458, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.023112788796424866, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.014534702524542809, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012210920453071594, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01190570741891861, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011005807667970657, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010754860937595367, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007591692265123129, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0077148377895355225, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0069926464930176735, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005213371943682432, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10354389250278473, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10354389250278473, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09319326281547546, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08582321554422379, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.08181510865688324, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07290741801261902, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04339883103966713, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04002608731389046, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05150148645043373, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04743451997637749, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04393433406949043, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.037743717432022095, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.0359470508992672, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.026139266788959503, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.022667501121759415, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.020804131403565407, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.020328111946582794, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.013042764738202095, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.0107091274112463, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.010359358973801136, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.00957943219691515, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.009278829209506512, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006778812501579523, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006722348276525736, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006135421805083752, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004323047585785389, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09319326281547546, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09319326281547546, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.23062309622764587, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.21511848270893097, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.2090659886598587, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.18726705014705658, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.10893169790506363, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.10298917442560196, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.12390882521867752, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11370620876550674, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.1100461557507515, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.09589781612157822, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.09098171442747116, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06311182677745819, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.05429593846201897, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.052149537950754166, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05160537734627724, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03148964047431946, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.026317035779356956, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.025826262310147285, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.023561686277389526, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.02322138100862503, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.016056835651397705, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.015289726667106152, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01512658130377531, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.00895326305180788, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11370620876550674, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11370620876550674, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.23423370718955994, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.21412000060081482, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.2065276950597763, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.17966803908348083, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10956846177577972, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10204701125621796, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12702491879463196, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11532618850469589, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11114949733018875, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09318970888853073, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08741850405931473, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06524629145860672, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05554085597395897, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.052925221621990204, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.052282772958278656, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03270934149622917, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.027845731005072594, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027299517765641212, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.024641036987304688, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02424938604235649, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01762174628674984, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017947202548384666, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016745883971452713, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.012906366027891636, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11532618850469589, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11532618850469589, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1836673468351364, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17248882353305817, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16851137578487396, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15232747793197632, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08712077885866165, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08305107057094574, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09712795913219452, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08992970734834671, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08793746680021286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07763803005218506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07386663556098938, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.049551822245121, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04316508024930954, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0419100821018219, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04161710664629936, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02478267066180706, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021810075268149376, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02152065560221672, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01991887204349041, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.019728822633624077, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013115552254021168, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013563796877861023, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012701783329248428, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009440099820494652, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09712795913219452, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09712795913219452, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.25907593965530396, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.24342535436153412, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2381073236465454, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.21551868319511414, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12288574129343033, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11729367822408676, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.13671420514583588, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12635710835456848, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12399027496576309, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10958223789930344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10421426594257355, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06957723200321198, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.060420386493206024, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0588233508169651, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05843552574515343, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.034735213965177536, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.029972907155752182, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.029575306922197342, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02722196839749813, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02697683684527874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.017975416034460068, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.017654653638601303, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.017412256449460983, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011095914989709854, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10958223789930344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10958223789930344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.27056851983070374, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.24321049451828003, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.23324443399906158, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.2069690078496933, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12538397312164307, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.11567390710115433, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14529691636562347, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.13367749750614166, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12920211255550385, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10841795057058334, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.10101388394832611, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07438350468873978, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06443121284246445, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.06065703183412552, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.059747036546468735, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.037327125668525696, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.03222567215561867, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.032083235681056976, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02896696887910366, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.028262676671147346, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.020227234810590744, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.021240249276161194, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.019033750519156456, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.015566961839795113, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10841795057058334, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10841795057058334, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10817279666662216, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10059468448162079, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09714192897081375, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08712441474199295, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.051095835864543915, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.047974247485399246, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05881928652524948, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05417634919285774, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.051693931221961975, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04491221904754639, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.042657408863306046, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.029952915385365486, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.025966545566916466, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02460514008998871, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02427147701382637, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.014982635155320168, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012797556817531586, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012526227161288261, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.0115587767213583, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011353696696460247, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007867901585996151, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008054858073592186, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.00738083990290761, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005516834557056427, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10817279666662216, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10817279666662216, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.0869065523147583, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08065084367990494, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.07701890170574188, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.06884875148534775, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04048972949385643, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.03753630816936493, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.04769021272659302, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.044332437217235565, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.040955424308776855, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.03555954247713089, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03379848971962929, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.024119023233652115, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.021110236644744873, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.019430799409747124, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.01899566315114498, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.012082531116902828, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.009948045946657658, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.009639791212975979, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.008945010602474213, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.008671686984598637, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006261234171688557, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006157052703201771, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.005739783402532339, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0038644508458673954, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.0869065523147583, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.0869065523147583, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.2592935562133789, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.24185173213481903, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.2358909547328949, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.21146155893802643, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12255315482616425, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.11610015481710434, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.13804474472999573, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.12666304409503937, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.12378479540348053, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10789605975151062, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.10250131785869598, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07040690630674362, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.06047545000910759, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.0585593655705452, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.0581413134932518, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03522513806819916, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.029589928686618805, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.02912057936191559, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.026515724137425423, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.026212550699710846, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.01797642558813095, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.017067410051822662, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.017070380970835686, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.01004987582564354, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10789605975151062, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10789605975151062, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.24215461313724518, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.22161777317523956, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.21380552649497986, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1903795450925827, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11238230019807816, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10449326038360596, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12902824580669403, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11893121153116226, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11426781862974167, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09736368060112, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.09242265671491623, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06625954806804657, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05736690014600754, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05433017387986183, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05358853191137314, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03327787294983864, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.028802920132875443, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028212251141667366, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025883827358484268, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.025426572188735008, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01801672764122486, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01887339912354946, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017079442739486694, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013822946697473526, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11893121153116226, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11893121153116226, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.18220782279968262, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17098255455493927, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16693654656410217, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15090060234069824, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0864524245262146, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08235178887844086, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09645909070968628, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.089301198720932, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.0872742235660553, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07695712149143219, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07320592552423477, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04920806735754013, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.042850978672504425, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04159076511859894, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04128619283437729, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.024608811363577843, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021658433601260185, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02136479690670967, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019758926704525948, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01956862583756447, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013048097491264343, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013486815616488457, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012629848904907703, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009395923465490341, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09645909070968628, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09645909070968628, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2570846974849701, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.24148188531398773, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.23613262176513672, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.21354445815086365, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12191261351108551, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11627837270498276, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1358468383550644, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12546581029891968, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12304001301527023, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10855695605278015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10325511544942856, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06910379230976105, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05997370183467865, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05837209150195122, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05799466371536255, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.034518610686063766, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02976377122104168, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.029360154643654823, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02699507400393486, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.026750344783067703, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.017948316410183907, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.017558325082063675, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.017388738691806793, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011072917841374874, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10855695605278015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10855695605278015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2727913558483124, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.24539479613304138, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.23526537418365479, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.20906171202659607, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12632493674755096, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.11661017686128616, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14695152640342712, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1350165754556656, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.13017913699150085, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10946830362081528, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.10207527130842209, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.0750863328576088, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06497818231582642, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.06100033223628998, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.0600559264421463, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03764408454298973, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.03223760426044464, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.032078519463539124, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.028971383348107338, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.028234368190169334, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.020258687436580658, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.021044369786977768, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01898077316582203, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.015160942450165749, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10946830362081528, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10946830362081528, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10111250728368759, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09422391653060913, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09099704027175903, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08165261894464493, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04784396290779114, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04493490979075432, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05521485209465027, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05078970640897751, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04836852476000786, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.042136870324611664, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.040124960243701935, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.028132202103734016, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.024336954578757286, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.022993696853518486, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.022666996344923973, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.014059926383197308, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.011873644776642323, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011607551947236061, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010717133060097694, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010507053695619106, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007320370990782976, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0073537565767765045, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006823268719017506, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0048612006939947605, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10111250728368759, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10111250728368759, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08814463019371033, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08193285763263702, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.0785025879740715, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07036745548248291, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04120943322777748, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.03837524354457855, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.048396605998277664, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04479173570871353, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04163181409239769, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.03631965070962906, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03445963189005852, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.02444387786090374, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.021359529346227646, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.019771933555603027, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.019386015832424164, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.012231497094035149, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.010089793242514133, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.009799059480428696, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009080926887691021, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.008829743601381779, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006339034531265497, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.00618712417781353, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.005820243153721094, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.003838206874206662, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08814463019371033, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08814463019371033, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.2546353340148926, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.23723505437374115, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.23104256391525269, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.2070387452840805, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12025155127048492, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.11385398358106613, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.13668741285800934, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.12466303259134293, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.12155953049659729, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10575010627508163, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.10041448473930359, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06967312842607498, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.059636592864990234, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05754788964986801, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.0570840947329998, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03478825092315674, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.029088608920574188, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.028607839718461037, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.026056531816720963, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.025729531422257423, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.01777811348438263, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.016904834657907486, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.016795393079519272, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.010010749101638794, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12025155127048492, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12025155127048492, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2305113673210144, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2075711190700531, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.19760535657405853, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.17673484981060028, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10644197463989258, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09683059900999069, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1262865513563156, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11580363661050797, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10885500907897949, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09129377454519272, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0873674750328064, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06447543203830719, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05587395653128624, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05166783183813095, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05062691122293472, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03261922299861908, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.027735775336623192, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02696656621992588, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.024868642911314964, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.024234304204583168, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01799527369439602, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018788546323776245, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016694312915205956, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014026879332959652, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11580363661050797, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11580363661050797, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1840808391571045, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17272911965847015, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16851826012134552, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15230073034763336, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0873633399605751, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08317737281322479, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.0977584570646286, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09047303348779678, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.0881916731595993, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07779289036989212, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07398909330368042, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04989675432443619, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04347461834549904, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04208916053175926, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04175284504890442, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.024965284392237663, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02201022207736969, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.021702179685235023, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.0201132632791996, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01991312950849533, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013270076364278793, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013849870301783085, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012818102724850178, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009800469502806664, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.0977584570646286, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.0977584570646286, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2481907308101654, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.23311099410057068, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.22781486809253693, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20597565174102783, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11767155677080154, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11217845976352692, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.13154277205467224, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12140852212905884, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11875894665718079, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10484180599451065, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09969447553157806, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06687907874584198, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05803940072655678, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.056367650628089905, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05597057193517685, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.033409491181373596, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.028826206922531128, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.028429539874196053, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02618389204144478, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02592640556395054, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.017371583729982376, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.017154233530163765, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016783908009529114, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010976346209645271, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12140852212905884, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11767155677080154, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.27196046710014343, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.24252110719680786, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2314111590385437, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.20496642589569092, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12583015859127045, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.11516772210597992, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14769691228866577, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.13551895320415497, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.13006260991096497, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10810007154941559, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.10084694623947144, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07571735233068466, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06531046330928802, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.06089038401842117, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05980195105075836, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03800787776708603, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.03228718414902687, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.032081637531518936, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.028854943811893463, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02804221585392952, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.020523859187960625, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.0212973915040493, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.019115429371595383, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.015438406728208065, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10810007154941559, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10810007154941559, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10654877871274948, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09930139780044556, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09551705420017242, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.0857313871383667, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0504273921251297, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04715107008814812, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05884118378162384, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.054196372628211975, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.050976648926734924, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04450523108243942, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04240047559142113, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.030050812289118767, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026089875027537346, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.024335656315088272, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02390315942466259, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015060847625136375, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.0127079077064991, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012384681962430477, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011521611362695694, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011250382289290428, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0079165268689394, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008132120594382286, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007300438825041056, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005612575449049473, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10654877871274948, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10654877871274948, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08538312464952469, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.07904203981161118, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.07508527487516403, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.06699524074792862, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.03973879665136337, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.036595419049263, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.04770831763744354, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04414159804582596, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04027092084288597, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.03491077944636345, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03317807614803314, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.024011965841054916, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.021071312949061394, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.01913636550307274, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.018636226654052734, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.012040491215884686, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.009874513372778893, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.00952537264674902, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.008891616016626358, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.008576851338148117, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006260857451707125, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006283114664256573, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.005637985188513994, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0040821232832968235, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08538312464952469, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08538312464952469, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.22805088758468628, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.21261391043663025, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.20619302988052368, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.1850138157606125, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.10745212435722351, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.10120221227407455, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.1237567663192749, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11274322867393494, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.10850459337234497, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.09468287229537964, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.09045466780662537, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06321610510349274, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.05399749428033829, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05140728875994682, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.050783123821020126, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03159372881054878, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.025990216061472893, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.025436988100409508, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.02331853099167347, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.022897828370332718, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.016116630285978317, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.015254991129040718, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.014973491430282593, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.00897612702101469, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11274322867393494, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11274322867393494, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22965237498283386, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20967069268226624, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.19730587303638458, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.17237652838230133, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10701164603233337, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09707167744636536, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13279491662979126, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.120976522564888, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10854131728410721, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09148801118135452, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08721313625574112, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06814736872911453, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05860951170325279, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05233946442604065, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05074894055724144, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03453303873538971, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02859361469745636, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027572721242904663, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025710556656122208, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02477177232503891, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.019161606207489967, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.020234176889061928, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017251238226890564, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01548596192151308, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.120976522564888, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.120976522564888, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.18031910061836243, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1692713499069214, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16508454084396362, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14926473796367645, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08552324026823044, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08140867203474045, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09590397030115128, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08867383003234863, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08632256835699081, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0762583538889885, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07254916429519653, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.048942066729068756, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04263431578874588, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04123527929186821, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.040902599692344666, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.024494854733347893, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02160360850393772, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02129744179546833, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01976647414267063, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.019559891894459724, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01306387409567833, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013668181374669075, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012602797709405422, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.0097264489158988, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09590397030115128, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09590397030115128, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23254600167274475, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2184232771396637, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2134186178445816, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1930743306875229, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11017199605703354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10500826686620712, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12309972941875458, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1137460395693779, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11115624755620956, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0982234999537468, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09340076148509979, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06261038035154343, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05437835305929184, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0527496263384819, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05236717313528061, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03125252574682236, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026964638382196426, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026589039713144302, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024505531415343285, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024264710023999214, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01617090217769146, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016042187809944153, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01559656485915184, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010237736627459526, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1137460395693779, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1137460395693779, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2661518156528473, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.23581808805465698, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2243819683790207, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1989479660987854, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12274676561355591, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.11188007891178131, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1449929028749466, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.13269582390785217, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.1271751970052719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10509918630123138, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09834118187427521, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07437021285295486, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06402123719453812, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05951451137661934, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05842047184705734, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03745146468281746, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.031814876943826675, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.03164134919643402, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.028456304222345352, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.027640175074338913, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.020465217530727386, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.021330511197447777, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.019020559266209602, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.015797536820173264, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10509918630123138, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10509918630123138, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11306379735469818, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10579120367765427, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1024746224284172, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09236577898263931, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05344946309924126, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.050441909581422806, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.061217501759529114, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05639849975705147, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0539589561522007, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.047413356602191925, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04514452815055847, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0311215128749609, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.027011003345251083, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0256536565721035, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02533215470612049, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015544474124908447, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013220812194049358, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01295448187738657, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011993820779025555, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011789172887802124, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008092744275927544, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008127061650156975, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0076204040087759495, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.00534807238727808, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11306379735469818, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11306379735469818, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.0834851861000061, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.07785433530807495, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.07445430010557175, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.0668388307094574, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.03895718604326248, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.036233674734830856, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.045865342020988464, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04256651550531387, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.03931944817304611, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.034444861114025116, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03277582675218582, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.023091677576303482, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.020301800221204758, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.018658602610230446, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.01827056333422661, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.011553221382200718, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.009543783031404018, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.00925152562558651, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.00862982776015997, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.008372287265956402, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.005991759710013866, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.005893114488571882, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.00549271097406745, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.003683371702209115, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.0834851861000061, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.0834851861000061, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.26033490896224976, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.2428041696548462, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.23629847168922424, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.21248216927051544, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12273699790239334, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.11603076010942459, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.13882672786712646, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.12758608162403107, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.12398988008499146, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.1084020733833313, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.10306525230407715, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07095026224851608, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.061084628105163574, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.058689460158348083, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05816107615828514, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.035465631633996964, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.029690198600292206, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.02915593795478344, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.02671111561357975, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.026313478127121925, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.01816551573574543, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.01733923703432083, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01713913306593895, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.0103018619120121, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.1084020733833313, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.1084020733833313, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.23772230744361877, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.21969565749168396, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.21406155824661255, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.18513503670692444, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11154398322105408, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10568094998598099, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12692973017692566, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11487041413784027, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11265421658754349, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09521203488111496, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08892307430505753, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06542328000068665, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.055514827370643616, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0539981871843338, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.053659841418266296, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.032911986112594604, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02865499071776867, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028287064284086227, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025470517575740814, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02524472586810589, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017863091081380844, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01853308081626892, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017396096140146255, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013750609941780567, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11487041413784027, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11487041413784027, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.18360252678394318, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17226940393447876, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1681852787733078, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15214481949806213, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0869520902633667, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.0828586146235466, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09709092974662781, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08988343179225922, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08774128556251526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07750549167394638, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07380792498588562, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04949214681982994, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04311515763401985, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04179072380065918, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04147860035300255, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02475328929722309, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02174154669046402, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.021438000723719597, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019855596125125885, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.019653556868433952, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013089718297123909, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013516202569007874, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012655937112867832, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009375534020364285, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09709092974662781, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09709092974662781, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.22997556626796722, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2160925269126892, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2112985998392105, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1912839710712433, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10886794328689575, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.1038842499256134, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12146499007940292, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11214277893304825, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10983351618051529, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0971384197473526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09243911504745483, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.061778318136930466, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05361027270555496, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.052119676023721695, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05176154524087906, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030885009095072746, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026645613834261894, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026288682594895363, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024236135184764862, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024005578830838203, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016113197430968285, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015840625390410423, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015604496002197266, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01013613399118185, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12146499007940292, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12146499007940292, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.26342716813087463, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.23440547287464142, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2233295887708664, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1985074132680893, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12138431519269943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.11082892119884491, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14301645755767822, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.13113215565681458, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12557056546211243, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.1044747605919838, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09778644144535065, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07306402176618576, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06317467987537384, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05874649062752724, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05767632648348808, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03676971420645714, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.031232496723532677, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.031026974320411682, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02798331528902054, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02717752754688263, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.020005296915769577, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.020720744505524635, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.018611304461956024, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.015102018602192402, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12138431519269943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.12138431519269943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11608075350522995, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.108706034719944, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10538692772388458, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09502118825912476, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05493522062897682, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05192602425813675, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06280674785375595, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05789909139275551, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.055498432368040085, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04874679818749428, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04642507806420326, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03194987028837204, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02771332859992981, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.026369962841272354, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.026048405095934868, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015950258821249008, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013589496724307537, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01332919578999281, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01233153697103262, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.012128883972764015, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.00831687729805708, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.00834373664110899, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007845600135624409, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005495818331837654, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11608075350522995, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11608075350522995, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08926033228635788, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08336768299341202, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.07957004010677338, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07157914340496063, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04172796383500099, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.03880762681365013, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.04914797097444534, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04579925164580345, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.042201172560453415, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.03700878471136093, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03524843230843544, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.024917857721447945, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.02189451828598976, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.020029906183481216, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.019561778753995895, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.01245820987969637, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.010258856229484081, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.009929925203323364, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009294351562857628, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.008995898067951202, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006496619898825884, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006385680288076401, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.0059270188212394714, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004006839357316494, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08926033228635788, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.08926033228635788, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.26869717240333557, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.2502693235874176, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.24326243996620178, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.2187175303697586, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12699374556541443, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.11978322267532349, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.14476710557937622, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.13267545402050018, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.12833616137504578, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11188416182994843, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.10695752501487732, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07414186000823975, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.06353883445262909, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.0606871135532856, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.059983525425195694, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03706229478120804, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.030648861080408096, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.030027924105525017, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.027522264048457146, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.027060680091381073, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.018909990787506104, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.01787957362830639, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01767951250076294, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.010476331226527691, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11188416182994843, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11188416182994843, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.23328769207000732, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2142457813024521, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.2045821249485016, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.18496806919574738, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10778432339429855, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09939749538898468, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1294572353363037, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11829843372106552, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10995782911777496, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09508039057254791, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.09153953939676285, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06618691235780716, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.056972187012434006, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05223187431693077, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05107488855719566, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03329487517476082, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02785634435713291, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02707059495151043, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025398798286914825, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.024689462035894394, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01816275715827942, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018740566447377205, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016670431941747665, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013733485713601112, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11829843372106552, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11829843372106552, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.18840888142585754, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17696616053581238, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1729481816291809, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15659312903881073, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08933819830417633, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08518263697624207, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09954197704792023, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09219477325677872, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09011498093605042, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07968931645154953, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07582017779350281, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05077319219708443, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04421817511320114, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04293625056743622, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.042635511606931686, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.025382185354828835, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02232307568192482, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.022024603560566902, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.020396092906594276, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.020206760615110397, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013429053127765656, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013852395117282867, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013011261820793152, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009598664939403534, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09954197704792023, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09954197704792023, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24724677205085754, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.23252438008785248, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2275320291519165, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.2061200886964798, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1170845627784729, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11184468120336533, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.13020724058151245, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12045620381832123, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11813132464885712, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10458147525787354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09958196431398392, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06625382602214813, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05755433440208435, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05604037269949913, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05567370727658272, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03307073935866356, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.028575308620929718, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02820546366274357, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025993268936872482, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.025760145857930183, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01713908091187477, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016866672784090042, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01662338525056839, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01064835861325264, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12045620381832123, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11813132464885712, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.261388897895813, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.23146803677082062, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2192091941833496, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19493016600608826, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11995511502027512, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10875431448221207, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14355801045894623, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.13151730597019196, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12454251945018768, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10304436087608337, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09668011963367462, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07323775440454483, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06323087960481644, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.058041684329509735, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.056778110563755035, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.036783237010240555, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.03083856962621212, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.030559901148080826, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.027583887800574303, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.026653535664081573, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019876427948474884, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.020547864958643913, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.018195854499936104, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014838241040706635, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11995511502027512, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11995511502027512, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11014960706233978, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10303682088851929, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09975264966487885, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08983810245990753, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05214208737015724, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04913147911429405, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.059965237975120544, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05520770326256752, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05267639458179474, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0462271049618721, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0440535694360733, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.030581269413232803, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026502462103962898, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025056837126612663, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.024709021672606468, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015290427021682262, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012951602227985859, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012676910497248173, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011755000799894333, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011533237062394619, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007989978417754173, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008040698245167732, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.00747841689735651, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005349848419427872, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11014960706233978, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11014960706233978, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09476891905069351, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.0883227214217186, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.08429363369941711, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.0757099986076355, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.0442793183028698, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04115474224090576, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05225217714905739, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.048655930906534195, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04475582391023636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.03908133506774902, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.037264008074998856, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.026369847357273102, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.02313285507261753, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.021212182939052582, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.020756185054779053, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.013158909045159817, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.01089986227452755, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.010564567521214485, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009838011115789413, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.009544393047690392, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006882904097437859, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006782528478652239, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.0062884376384317875, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004286787938326597, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09476891905069351, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09476891905069351, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.25928956270217896, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.24224399030208588, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.23580721020698547, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.21220609545707703, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12263990193605423, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.11605669558048248, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.13958021998405457, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.12757804989814758, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.12389341741800308, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10837080329656601, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.1034700945019722, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07127653807401657, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.06097486987709999, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05858835577964783, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05802549421787262, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03559672087430954, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.0296078622341156, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.029062369838356972, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.026615723967552185, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.026251234114170074, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.01815900392830372, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.017171433195471764, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.017047356814146042, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.010077917017042637, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10837080329656601, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10837080329656601, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.21928785741329193, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20128397643566132, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.19343701004981995, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1718589961528778, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.1017158105969429, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09413722902536392, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1188013032078743, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10906998813152313, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10345394164323807, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08828695118427277, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08341935276985168, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.060799770057201385, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05264098197221756, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04923086240887642, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04840872809290886, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.030594054609537125, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02626461163163185, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02561882510781288, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.023667342960834503, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02316822111606598, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01659797504544258, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01748962141573429, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015540768392384052, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.012930777855217457, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1188013032078743, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1188013032078743, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1924171894788742, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18078497052192688, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.17657627165317535, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1598728895187378, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0910847932100296, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08683928102254868, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10165759176015854, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09402233362197876, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09191589057445526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08130169659852982, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07741611450910568, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.051739949733018875, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04503079131245613, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.043719206005334854, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04340362548828125, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.025888126343488693, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.022631263360381126, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.022327423095703125, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.020667795091867447, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.020469732582569122, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013686527498066425, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013903163373470306, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013251269236207008, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009461373090744019, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10165759176015854, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10165759176015854, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24262812733650208, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22825324535369873, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.22318798303604126, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.2022133618593216, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11487086117267609, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10968350619077682, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12797856330871582, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11822834610939026, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11593106389045715, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10258878022432327, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09773349016904831, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0651298314332962, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05651019886136055, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05499841272830963, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05463607609272003, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.0325707383453846, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.028080971911549568, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02771170809864998, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025549525395035744, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.025323396548628807, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01701761595904827, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01663188263773918, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016495611518621445, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01057179644703865, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11822834610939026, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11822834610939026, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2558002471923828, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2251521199941635, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2125939130783081, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1891299933195114, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11710931360721588, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.1055966168642044, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14027932286262512, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12896686792373657, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12188348919153214, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10030367970466614, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09408961236476898, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07172815501689911, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06208828464150429, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.056675802916288376, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.0553477518260479, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.036114804446697235, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.030109724029898643, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029816316440701485, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026867743581533432, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025901472195982933, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01954837143421173, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.020114997401833534, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01782902516424656, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014472863636910915, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11710931360721588, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11710931360721588, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11110243201255798, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10409969836473465, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10084173828363419, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09111185371875763, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05262034386396408, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04967721924185753, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06023944914340973, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05558182671666145, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05314087122678757, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04676879197359085, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.044625796377658844, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.030681541189551353, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026626231148838997, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025261888280510902, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02493252232670784, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01533588394522667, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013006751425564289, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012743642553687096, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011814854107797146, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011605842970311642, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007986726239323616, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007986193522810936, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007514507509768009, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005230833310633898, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11110243201255798, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11110243201255798, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09363710880279541, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08739571273326874, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.08372346311807632, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07535402476787567, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.043873537331819534, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.0409027598798275, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.051332347095012665, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04785284027457237, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.044371265918016434, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.038904592394828796, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03698118403553963, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.026010742411017418, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.022843055427074432, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.021037768572568893, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.02059047669172287, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.012986191548407078, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.010781636461615562, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.010456312447786331, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009760349057614803, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.00946823414415121, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006755076348781586, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006670642644166946, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.00620255945250392, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004181407857686281, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09363710880279541, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09363710880279541, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.27837228775024414, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.26042407751083374, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.2537209689617157, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.22875554859638214, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.13188326358795166, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.1250280886888504, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.14907938241958618, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.13693076372146606, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.13323916494846344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11672341078519821, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.11120659857988358, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07603035867214203, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.06542988121509552, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.0631185844540596, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.06262214481830597, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.0380057692527771, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.03183833137154579, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.03130011260509491, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.028655488044023514, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.028287380933761597, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.01936524733901024, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.018420888110995293, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.018336251378059387, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.010807509534060955, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11672341078519821, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11672341078519821, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1809326410293579, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.15813671052455902, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1486254781484604, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1287214607000351, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0840136855840683, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07501904666423798, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10021818429231644, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09184042364358902, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08568239957094193, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06841600686311722, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06325627863407135, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05161849781870842, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.044530294835567474, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04097563773393631, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04012555629014969, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02616814710199833, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.022311696782708168, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.021622508764266968, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01941666007041931, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01885424554347992, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01471226941794157, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.015484750270843506, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.013658386655151844, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01187003031373024, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10021818429231644, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10021818429231644, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.19590722024440765, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18415535986423492, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18006117641925812, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16305308043956757, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09261878579854965, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08838684111833572, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.103166364133358, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09552588313817978, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09346731752157211, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08279424160718918, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07888868451118469, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05253985896706581, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04573700577020645, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04442519694566727, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04411549121141434, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.026254668831825256, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.022926371544599533, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02262575738132, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.020945146679878235, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.020746195688843727, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013783213682472706, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01398166548460722, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013352220878005028, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00940695870667696, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.103166364133358, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.103166364133358, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23618078231811523, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2222551852464676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21745991706848145, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19710493087768555, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11171876639127731, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10673729330301285, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.124420665204525, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11495858430862427, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11270836740732193, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09994008392095566, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09521842002868652, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06319984793663025, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.054908812046051025, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.053449369966983795, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05309437960386276, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.031567104160785675, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.027261968702077866, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026914935559034348, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024828607216477394, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024605643004179, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01641690731048584, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01610986515879631, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015909520909190178, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010195432230830193, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11495858430862427, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11495858430862427, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2539728581905365, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22414065897464752, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21167951822280884, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18785902857780457, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11615443974733353, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.1050017923116684, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.14058636128902435, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12825284898281097, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12090107053518295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09980297833681107, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09344689548015594, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07162075489759445, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06177503988146782, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.056275516748428345, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.054929714649915695, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03606612607836723, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.03000582754611969, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.02970004454255104, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026872163638472557, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025905923917889595, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019692644476890564, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.020194372162222862, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01790996640920639, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014680488035082817, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12090107053518295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12090107053518295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.22.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11403417587280273, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1067904382944107, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10350558161735535, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.0935012549161911, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05393834039568901, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05094636604189873, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06180773675441742, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05689816176891327, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0544712096452713, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.047974202781915665, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04574989899992943, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.031481239944696426, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.027260197326540947, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025922205299139023, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.025595197454094887, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015747398138046265, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013381125405430794, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.013114229775965214, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.012167935259640217, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01196207944303751, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008214757777750492, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008264805190265179, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007721406407654285, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005488233175128698, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11403417587280273, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11403417587280273, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.22.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.0952269434928894, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.0888146162033081, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.08485172688961029, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07629094272851944, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.044587425887584686, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04137728735804558, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05248815566301346, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04882344976067543, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04505797103047371, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.039466582238674164, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.037586648017168045, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.026522397994995117, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.023353058844804764, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.021336443722248077, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.020847810432314873, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.013252967037260532, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.010899168439209461, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.010556070134043694, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009853549301624298, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.00955017376691103, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006860025227069855, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006720059085637331, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006264599971473217, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004151395987719297, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.0952269434928894, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.0952269434928894, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.22.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.2549070119857788, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.23782309889793396, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.23115213215351105, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.20786666870117188, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12018585950136185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.11348909884691238, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.13651125133037567, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1254683881998062, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.12146445363759995, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10616572946310043, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.10127444565296173, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.0696294829249382, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.059981219470500946, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05749804899096489, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.056896768510341644, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03478189557790756, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.029076244682073593, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.028535014018416405, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.02614332176744938, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.025743603706359863, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.0177957471460104, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.016972046345472336, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01672685518860817, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.010051167570054531, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.12146445363759995, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12018585950136185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.22.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.24217388033866882, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2185772955417633, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.21010175347328186, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.18376345932483673, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11251398921012878, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10394180566072464, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1298341155052185, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11886086314916611, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11429879069328308, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09465838968753815, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08917582780122757, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06667723506689072, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.057455874979496, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05451555177569389, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05381401628255844, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03361078351736069, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02916964516043663, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028574872761964798, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025761203840374947, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.025321858003735542, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01852772943675518, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019382750615477562, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017627574503421783, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014511818066239357, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11886086314916611, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11886086314916611, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.22.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.19284962117671967, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18137849867343903, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.17731793224811554, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16081418097019196, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09121913462877274, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08710399270057678, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10183192044496536, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09409476816654205, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09203565120697021, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08160112053155899, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0779014304280281, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05186909809708595, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04507644847035408, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.043791260570287704, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.0434931255877018, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.026003360748291016, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.022703474387526512, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02240716852247715, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.020774664357304573, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.020578861236572266, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013851463794708252, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013978928327560425, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013434226624667645, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009562993422150612, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10183192044496536, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10183192044496536, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.22.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23107671737670898, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21748575568199158, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21272709965705872, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19292835891246796, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10923592001199722, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10437528789043427, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12165949493646622, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11237134784460068, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11019027233123779, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09769327193498611, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09320694953203201, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06191225349903107, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05370152369141579, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.052270013839006424, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.051932185888290405, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030912648886442184, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02668776735663414, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.0263445395976305, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024321677163243294, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02410249598324299, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.0161259975284338, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015813691541552544, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01564045064151287, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010066409595310688, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11237134784460068, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11237134784460068, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.22.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24171948432922363, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.21352043747901917, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20148834586143494, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.17780078947544098, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11085216701030731, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10001099109649658, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13338616490364075, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12234485894441605, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11513915657997131, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09479965269565582, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08843725919723511, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06847433000802994, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.0589880533516407, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05365487188100815, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05234536901116371, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03456630930304527, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.028525572270154953, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028217535465955734, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.025442836806178093, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.024501577019691467, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01896301656961441, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01912819594144821, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017257047817111015, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013808880932629108, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11513915657997131, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11513915657997131, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11165835708379745, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10446471720933914, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10100489854812622, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09110084921121597, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05279263108968735, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04969129338860512, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.060928404331207275, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05610070750117302, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.053336143493652344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04684271663427353, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04471856355667114, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.030983073636889458, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026838375255465508, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025358200073242188, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.024997755885124207, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015481011010706425, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013084622099995613, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012796944007277489, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011881315149366856, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011654670350253582, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008065521717071533, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008097438141703606, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007529538590461016, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005349544808268547, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11165835708379745, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11165835708379745, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09314389526844025, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08668171614408493, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.08249570429325104, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07413218915462494, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.043458305299282074, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04023640602827072, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.051597557961940765, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04797939956188202, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.044014766812324524, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.0384373664855957, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03659326210618019, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.026040835306048393, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.022939275950193405, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.02087397128343582, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.020384062081575394, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.013041814789175987, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.01069874968379736, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.01032157801091671, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009664392098784447, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.009329783730208874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006760870572179556, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.0066675846464931965, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006132096983492374, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004147515632212162, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09314389526844025, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09314389526844025, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.2405540943145752, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.22177234292030334, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.21415932476520538, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.19131797552108765, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.11249173432588577, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.10498207062482834, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.12941694259643555, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11899607628583908, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11392711848020554, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.09811240434646606, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.09349490702152252, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06630273163318634, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.05693184584379196, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.053760819137096405, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05300251394510269, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03316290304064751, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.027168579399585724, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.02652628906071186, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.024184977635741234, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.023663081228733063, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.0168851837515831, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.016020001843571663, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.0155583331361413, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.009393738582730293, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11899607628583908, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.11899607628583908, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.18316088616847992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.16550660133361816, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.15973015129566193, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.137624591588974, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0849338248372078, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07838541269302368, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09682390093803406, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.08870916813611984, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08631589263677597, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07112104445695877, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06604772061109543, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04973427951335907, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04286430776119232, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.041174404323101044, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04077431559562683, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.024978457018733025, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.022034240886569023, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02161109820008278, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.019411373883485794, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.019161012023687363, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.013725358992815018, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.014628448523581028, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.013210546225309372, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.011053289286792278, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09682390093803406, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09682390093803406, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1979304850101471, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18629620969295502, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18229572474956512, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16539855301380157, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0935993641614914, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08944031596183777, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10411103069782257, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09641207009553909, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09442346543073654, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08381300419569016, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07988516241312027, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.052997417747974396, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04613090679049492, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04486784711480141, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04456951096653938, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.026467738673090935, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.023076649755239487, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02278009243309498, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.021082647144794464, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.020893169566988945, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013861782848834991, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013929254375398159, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013441601768136024, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009213516488671303, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10411103069782257, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10411103069782257, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23229607939720154, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21878495812416077, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21419477462768555, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.194339781999588, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1097479835152626, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10499372333288193, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.122163325548172, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11281891912221909, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11070885509252548, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09830605238676071, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09381193667650223, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06205027177929878, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.053904056549072266, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05252304673194885, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05220355838537216, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.031003417447209358, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026819344609975815, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026481425389647484, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024465473368763924, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024256817996501923, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01617085561156273, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015879372134804726, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01569848321378231, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010114620439708233, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11281891912221909, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11281891912221909, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24135398864746094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.21320974826812744, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20198741555213928, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.17782263457775116, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11061527580022812, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10021601617336273, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13193082809448242, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12085677683353424, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11476334184408188, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09423748403787613, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.0876198336482048, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.067434161901474, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05809292942285538, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.053474027663469315, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05232381448149681, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.034025296568870544, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.028278404846787453, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028039151802659035, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02510332129895687, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.024266913533210754, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01849040761590004, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01867501437664032, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.016969604417681694, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013353876769542694, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12085677683353424, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12085677683353424, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1085706576704979, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10163728892803192, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09830481559038162, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08885159343481064, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0513431616127491, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.048383504152297974, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.059120140969753265, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.054412808269262314, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.051878586411476135, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.045621976256370544, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.043537240475416183, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03007286973297596, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02603708952665329, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.024651750922203064, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02432151511311531, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015022079460322857, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012691215611994267, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012418996542692184, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011523965746164322, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011309009045362473, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.00782601349055767, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007798957172781229, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007335114758461714, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0050993748009204865, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1085706576704979, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1085706576704979, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09293967485427856, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08626274019479752, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.08208395540714264, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07383310049772263, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04337076097726822, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.040096715092659, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05145851522684097, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.04789356142282486, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04389224946498871, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.038233209401369095, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03641323745250702, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.02593962661921978, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.02277499809861183, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.02077549695968628, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.020271578803658485, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.01300373300909996, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.010608512908220291, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.010257012210786343, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009566930122673512, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.009256954304873943, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.00674700690433383, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.006573937367647886, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006145049352198839, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004073425196111202, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09293967485427856, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09293967485427856, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.24277257919311523, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.22491613030433655, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.21790803968906403, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.1958121955394745, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.11388823390007019, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.10683871060609818, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.13087984919548035, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1197613924741745, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11522188782691956, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10007236152887344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.09531880170106888, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.0667535588145256, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.057273998856544495, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05439399555325508, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05367480590939522, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.033249806612730026, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.02754543535411358, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.026933835819363594, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.024657797068357468, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.024202708154916763, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.016997674480080605, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.016217753291130066, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01579897105693817, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.009642851538956165, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1197613924741745, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1197613924741745, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.23403629660606384, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.21574871242046356, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.20954908430576324, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.18249543011188507, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10900037735700607, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10243849456310272, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12390584498643875, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11358314007520676, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1105223298072815, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0937311053276062, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08743268251419067, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06363580375909805, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.055025648325681686, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.052939921617507935, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05245227366685867, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.032004110515117645, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.028547378256917, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028075914829969406, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025659585371613503, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.025343772023916245, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017793938517570496, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019166866317391396, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017164407297968864, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01472178753465414, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11358314007520676, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11358314007520676, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.20360074937343597, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.19171112775802612, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18755079805850983, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.17015595734119415, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0962795689702034, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09198597073554993, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10716156661510468, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09915490448474884, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.097133569419384, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08621764928102493, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08224222809076309, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05449783802032471, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04739634320139885, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04609425365924835, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04579169303178787, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.027230601757764816, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.0236409492790699, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.023335525766015053, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02158668078482151, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.021387778222560883, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014244500547647476, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014167978428304195, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013813312165439129, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009232241660356522, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10716156661510468, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10716156661510468, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23404762148857117, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22047321498394012, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21585175395011902, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19592657685279846, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11066664010286331, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10584031045436859, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12285909056663513, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11365462839603424, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11161664873361588, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09911097586154938, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09459217637777328, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06245272606611252, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.054337769746780396, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05297331511974335, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05264715477824211, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03125341981649399, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02706635557115078, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026732146739959717, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02469875104725361, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02448856830596924, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016356170177459717, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01605166681110859, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015904342755675316, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010261362418532372, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11365462839603424, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11365462839603424, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.23139405250549316, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.20393165946006775, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.19280171394348145, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.16982991993427277, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.10572183132171631, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.09554261714220047, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.12694182991981506, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11617368459701538, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.10987585783004761, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09015917032957077, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08383259922266006, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.0645093321800232, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.0557699054479599, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05101644992828369, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.049853868782520294, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03233390301465988, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02691732719540596, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.02666415087878704, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.023899322375655174, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.023034365847706795, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.017344165593385696, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.017747581005096436, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.015797298401594162, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012545310892164707, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11617368459701538, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11617368459701538, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10861224681138992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10129769146442413, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09771580994129181, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08804424107074738, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05122395604848862, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04807977378368378, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05927001312375069, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.054630063474178314, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05180172994732857, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04530588537454605, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04320411756634712, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.030108503997325897, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02609914541244507, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02458932064473629, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.024229882284998894, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015062552876770496, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01263350062072277, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01233507040888071, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01142138335853815, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011184869334101677, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.00785665214061737, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007755037397146225, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007341369520872831, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.00500124366953969, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10861224681138992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10861224681138992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09297950565814972, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.08628645539283752, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.08213059604167938, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07366211712360382, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.043431103229522705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04003092274069786, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05167323350906372, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.048015885055065155, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.043948233127593994, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.038302771747112274, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.036484986543655396, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.026199111714959145, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.023000910878181458, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.020840955898165703, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.020298004150390625, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.013126635923981667, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.010721521452069283, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.010362387634813786, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.009672029875218868, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.009357242844998837, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.006893529556691647, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.0067563410848379135, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006265098694711924, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004345831461250782, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09297950565814972, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09297950565814972, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.2570837140083313, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.23724676668643951, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.22921572625637054, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.20499224960803986, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.1204194724559784, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.11240452527999878, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.13810518383979797, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1268768310546875, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.12192939966917038, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10511066764593124, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.09993046522140503, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.0705559104681015, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.060816869139671326, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05755648389458656, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05677194148302078, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03534076362848282, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.029124530032277107, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.02841041050851345, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.025914646685123444, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.025388268753886223, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.01803872175514698, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.017199525609612465, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.016664348542690277, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.010183035396039486, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.1204194724559784, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.1204194724559784, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.20592471957206726, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.16950207948684692, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1571529507637024, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1446210741996765, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09484893083572388, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08235950022935867, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11086428910493851, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10214310884475708, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09734532982110977, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07510259747505188, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07196646183729172, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05702735111117363, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.049838531762361526, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04653659090399742, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.045744262635707855, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02881496585905552, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.025791464373469353, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.025052834302186966, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02223057858645916, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.021731190383434296, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.016283005475997925, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01827894151210785, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015365137718617916, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01458627637475729, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11086428910493851, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11086428910493851, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.20895366370677948, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.19674140214920044, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.19251348078250885, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.17475317418575287, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0988534465432167, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09444792568683624, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10983659327030182, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10176746547222137, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09972183406352997, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08846915513277054, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08437217026948929, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05586428940296173, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04863669350743294, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04733077436685562, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.047017768025398254, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02790645696222782, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.024268563836812973, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.023957055062055588, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.022148745134472847, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02195395901799202, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014576492831110954, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014546212740242481, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01414666511118412, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009485399350523949, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10983659327030182, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10983659327030182, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2375299036502838, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22376702725887299, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21912389993667603, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19885189831256866, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11229342967271805, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10737846791744232, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12447060644626617, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1153317391872406, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11326207965612411, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10054565966129303, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09586445987224579, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06323644518852234, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05507281422615051, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.053675323724746704, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.053349755704402924, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.031551603227853775, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02730422280728817, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02696429006755352, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.0248829647898674, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024668507277965546, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016257943585515022, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01599831134080887, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01579144038259983, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009955823421478271, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1153317391872406, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1153317391872406, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2366725653409958, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.20950698852539062, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.19868795573711395, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.17561490833759308, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.1080239787697792, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.09818573296070099, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.12907926738262177, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11811626702547073, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11219814419746399, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09267033636569977, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08635887503623962, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06556500494480133, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.056590769439935684, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05202954635024071, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.0509074367582798, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03304445743560791, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.027245260775089264, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.027010995894670486, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02422296814620495, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.023382727056741714, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01798117347061634, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.017652172595262527, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.016537981107831, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.01217629387974739, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11811626702547073, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11811626702547073, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10786353051662445, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1010407879948616, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09784767031669617, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08850390464067459, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05098435655236244, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.048128627240657806, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05856706202030182, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05396556854248047, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.051497966051101685, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04534168541431427, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.043278396129608154, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.02969958819448948, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.025758076459169388, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.024460555985569954, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.024152010679244995, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.014832289889454842, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012567142024636269, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01231022085994482, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011414515785872936, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011211135424673557, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007730166893452406, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007677758112549782, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007284379564225674, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004971436690539122, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10786353051662445, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10786353051662445, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.1032770574092865, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.09646089375019073, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.09234318137168884, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.0832984447479248, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04846986383199692, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04520265385508537, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05685561150312424, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.052748024463653564, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04903081804513931, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.04300122335553169, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.04106350243091583, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.028775690123438835, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.02514297142624855, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.023232556879520416, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.022775748744606972, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.014360370114445686, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.011885982938110828, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.011533690616488457, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.010753919370472431, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.010452340357005596, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.0074773384258151054, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.007294547278434038, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006881407927721739, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004515211097896099, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.1032770574092865, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.1032770574092865, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.27283719182014465, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.25378453731536865, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.24674391746520996, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.22166883945465088, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.12851320207118988, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.12121312320232391, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.14570990204811096, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1338733583688736, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.13001032173633575, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11311156302690506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.10782160609960556, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07430887222290039, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.0640048012137413, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.061452716588974, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.060842808336019516, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.037190839648246765, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.031122246757149696, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.030549537390470505, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.027891699224710464, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.027513237670063972, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.01900199055671692, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.018240882083773613, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.017882078886032104, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.01091412827372551, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11311156302690506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11311156302690506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.200792133808136, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.18106473982334137, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.17353609204292297, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.14975181221961975, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09311887621879578, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08562486618757248, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10792627185583115, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09928996115922928, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09489098936319351, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07839391380548477, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07301497459411621, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.055269304662942886, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04800839722156525, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.045256249606609344, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.044581878930330276, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.027802282944321632, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02445116825401783, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.023908570408821106, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.021696865558624268, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.021287258714437485, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015281109139323235, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016608595848083496, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014448264613747597, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.012714982032775879, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10792627185583115, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10792627185583115, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.21575556695461273, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.20320697128772736, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1988653689622879, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.180541530251503, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10216973721981049, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09759803861379623, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11354772746562958, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10513406246900558, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10304136574268341, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09144221246242523, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08721961081027985, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05783160775899887, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.050283849239349365, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0489443764090538, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04862673580646515, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02888745442032814, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.025167761370539665, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.024849919602274895, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02299102582037449, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02279198169708252, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01516411267220974, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01517993863672018, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.014726330526173115, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010029764845967293, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11354772746562958, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11354772746562958, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24038419127464294, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2264154553413391, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2217765748500824, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20134159922599792, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11365620046854019, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10869515687227249, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12605318427085876, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1167132779955864, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11466765403747559, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10177591443061829, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09706693142652512, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0640442743897438, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05573025345802307, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.054352056235075, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05402683466672897, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03195766732096672, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.027626169845461845, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.027280595153570175, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025164447724819183, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024952635169029236, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016499334946274757, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01616641692817211, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01603492721915245, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010039175860583782, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1167132779955864, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1167132779955864, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.23530074954032898, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.20874914526939392, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.19752003252506256, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1749504804611206, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.10743295401334763, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.0974641963839531, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.12910017371177673, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11861782521009445, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11158761382102966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.0926252007484436, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.0863977000117302, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06557273864746094, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.056863438338041306, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05184754729270935, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05060488358139992, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03291357681155205, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02727714367210865, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.026957012712955475, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.024379095062613487, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02345374785363674, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.017597975209355354, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.017913559451699257, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01597486063838005, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012470013462007046, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11861782521009445, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11861782521009445, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11739787459373474, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1093529537320137, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10553708672523499, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09514733403921127, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05546310171484947, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05199837684631348, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06393957138061523, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05904656648635864, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.056098856031894684, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04899990186095238, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.046742238104343414, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03253254294395447, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.028274912387132645, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.026618661358952522, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.026221638545393944, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.016273919492959976, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013667390681803226, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.013344227336347103, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.0123421186581254, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.012091371230781078, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008449047803878784, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008377611637115479, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007883330807089806, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005395964719355106, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11739787459373474, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11739787459373474, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.10359002649784088, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.0957912877202034, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.09099068492650986, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.08168377727270126, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04853963479399681, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04456108808517456, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05811506509780884, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.053844571113586426, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04911995306611061, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.04264114052057266, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.04070934280753136, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.02941207028925419, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.025790084153413773, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.023250948637723923, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.022636504843831062, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.014666041359305382, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.011918595992028713, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.011465548537671566, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.010730472393333912, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.01031651720404625, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.007652482483536005, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.0074855005368590355, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006877135019749403, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004628748167306185, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.10359002649784088, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.10359002649784088, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.27980825304985046, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.25855860114097595, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.2503129541873932, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.2240312397480011, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.13143832981586456, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.12325213104486465, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.15081281960010529, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.138030007481575, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.1332189440727234, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11488338559865952, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.10939337313175201, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07700102776288986, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.06606506556272507, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.06299547851085663, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.06223927438259125, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.03852624446153641, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.03185761347413063, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.03118387795984745, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.028393490239977837, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.027887050062417984, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.019665297120809555, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.018769364804029465, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01835053786635399, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.011184653267264366, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11488338559865952, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11488338559865952, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.18855348229408264, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.16729265451431274, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1604154407978058, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.14115551114082336, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08766797930002213, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08014948666095734, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.100211001932621, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.091807521879673, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08928428590297699, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07194072008132935, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06918036192655563, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05139675363898277, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04461931809782982, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.042734548449516296, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04228623956441879, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02592952363193035, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.023301923647522926, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02285095304250717, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.020378468558192253, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.020107485353946686, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.014520281925797462, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.015964968129992485, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01396208442747593, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01253555342555046, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.100211001932621, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.100211001932621, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.22323943674564362, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21007557213306427, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2055959403514862, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18643994629383087, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10566230118274689, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10089316219091415, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1173747256398201, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10871624946594238, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10658521950244904, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09447207301855087, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08996666222810745, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.059672653675079346, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05195405334234238, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05056285113096237, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05023927241563797, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02979593351483345, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.025840427726507187, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.025507288053631783, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02353823184967041, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.023326396942138672, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01549032237380743, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01534254103899002, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015034771524369717, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009811545722186565, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1173747256398201, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1173747256398201, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24779435992240906, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2333039492368698, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.22846975922584534, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.2072691172361374, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1172453984618187, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11209762841463089, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12992508709430695, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12036900967359543, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11827024817466736, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10491020232439041, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0999096930027008, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06603895872831345, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05748594179749489, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.056053463369607925, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05572572350502014, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03294418752193451, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.028472013771533966, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02811560407280922, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025907237082719803, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02568943426012993, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016963250935077667, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016617998480796814, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01648106426000595, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010255258530378342, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12036900967359543, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11827024817466736, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.25107866525650024, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22492782771587372, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2137727439403534, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19000673294067383, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11504033207893372, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10521350055932999, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13798275589942932, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12603887915611267, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11895972490310669, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09986884146928787, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09333968907594681, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06965908408164978, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.060230553150177, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.055305566638708115, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05407541245222092, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.035007935017347336, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.028719669207930565, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028405606746673584, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.025691095739603043, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02478714846074581, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01868733949959278, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01832152158021927, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01706647127866745, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012244926765561104, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11895972490310669, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11895972490310669, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12027505785226822, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11198510229587555, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10809361189603806, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09742816537618637, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05689902976155281, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.053275495767593384, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06559566408395767, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06055331975221634, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0575542114675045, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.050196047872304916, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04784258455038071, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03338005021214485, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.029049761593341827, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0273154117166996, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.026906628161668777, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.016703374683856964, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014069712720811367, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.013731532730162144, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.012702158652245998, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.012433608993887901, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008695350028574467, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008684631437063217, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008104710839688778, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005664370488375425, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12027505785226822, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12027505785226822, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.1036551371216774, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.09602518379688263, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.0912201851606369, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.08185166120529175, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04853089526295662, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04466002807021141, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.0579291470348835, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.053751446306705475, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04917262867093086, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.04267481341958046, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.040710605680942535, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.02934195287525654, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.02571728453040123, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.0232954490929842, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.022698968648910522, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.014682884328067303, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.01193163264542818, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.01148600410670042, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.010719257406890392, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.010337580926716328, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.007629623170942068, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.007452426012605429, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006875535938888788, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004632852505892515, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.1036551371216774, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.1036551371216774, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.28702446818351746, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.2654138505458832, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.2570708394050598, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.23019835352897644, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.13514862954616547, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.1265326589345932, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.1546151340007782, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1417628526687622, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.13685311377048492, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11796286702156067, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.11223568767309189, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07883621752262115, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.06779049336910248, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.06459344923496246, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.0638294443488121, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.039405304938554764, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.032686181366443634, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.03200206533074379, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.029139444231987, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.028614966198801994, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.020127346739172935, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.019205594435334206, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01879533752799034, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.011345121078193188, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11796286702156067, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11796286702156067, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.14291217923164368, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.12374390661716461, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11580924689769745, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1010097935795784, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06563586741685867, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05758940055966377, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.07897986471652985, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.0722835510969162, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06781331449747086, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05414298549294472, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.050066377967596054, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04060342162847519, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.035263076424598694, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03228361904621124, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.031548116356134415, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.020551908761262894, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01793944649398327, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01737184077501297, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.015878185629844666, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.015426469966769218, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.011572729796171188, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.012902405112981796, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0106594767421484, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010282373987138271, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11580924689769745, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, "best_option": { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11580924689769745, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.28.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.22450943291187286, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21113868057727814, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.20647858083248138, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18718235194683075, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1063859686255455, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10151267051696777, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11843837052583694, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10956118255853653, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10734386742115021, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0950063094496727, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09056601673364639, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06026002764701843, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.052406542003154755, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0509473942220211, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05061342939734459, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03014404885470867, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02610575221478939, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02575690485537052, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.023768983781337738, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.023548582568764687, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015807300806045532, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015603767707943916, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015334081836044788, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010100656189024448, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11843837052583694, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11843837052583694, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.28.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24397586286067963, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22956453263759613, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.22464436292648315, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20372870564460754, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11553201824426651, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11036490648984909, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12840934097766876, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11878617107868195, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11659401655197144, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.1032457947731018, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0983169674873352, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06523139029741287, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.056743789464235306, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05527810752391815, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05493355914950371, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.032582879066467285, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02815423160791397, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.027794431895017624, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025607993826270103, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.025384677574038506, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016893811523914337, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01655159331858158, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01639348827302456, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010378491133451462, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11878617107868195, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11878617107868195, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.28.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2512762248516083, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22606782615184784, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2150501161813736, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1915084272623062, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11540006101131439, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10573244839906693, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1385573148727417, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12651704251766205, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11922929435968399, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10071857273578644, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09428112953901291, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07036971300840378, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06068205088376999, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05554927513003349, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05429806560277939, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03550688549876213, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02898469753563404, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028621092438697815, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02606872282922268, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025126179680228233, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019343385472893715, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.018681932240724564, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017672352492809296, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012664347887039185, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11922929435968399, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11922929435968399, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11222615838050842, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10449931025505066, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10079417377710342, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09079480171203613, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05310392007231712, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04974565654993057, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06144746392965317, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.056652940809726715, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.053722962737083435, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04682547226548195, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.044687628746032715, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.031259164214134216, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.027146825566887856, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025508293882012367, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.025122039020061493, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015643557533621788, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013128253631293774, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012812593951821327, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011848579160869122, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011595703661441803, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008171644061803818, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008093343116343021, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007637208793312311, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005263732746243477, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11222615838050842, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11222615838050842, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09731374680995941, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.09007878601551056, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.08559734374284744, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.07688257098197937, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04567525163292885, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04199240729212761, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05460715293884277, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.050612710416316986, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04625270888209343, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.04008496552705765, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.03816305845975876, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.027579806745052338, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.02419317327439785, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.021886948496103287, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.02134244330227375, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.013758246786892414, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.01125063095241785, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.010856030508875847, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.010102412663400173, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.009759679436683655, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.007182771805673838, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.007057573180645704, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006514081731438637, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004485566169023514, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09731374680995941, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.09731374680995941, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.28446662425994873, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.2627583146095276, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.25433117151260376, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.22774510085582733, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.13408005237579346, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.12546750903129578, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.15384717285633087, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.140965536236763, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.13590103387832642, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.1168556958436966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.11120016127824783, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07830590754747391, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.06745877861976624, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.06420189142227173, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.06339717656373978, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.039102114737033844, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.032527290284633636, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.03181621432304382, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.028953872621059418, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.02843485400080681, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.020011764019727707, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.019165873527526855, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01869523897767067, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.011502579785883427, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.1168556958436966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.1168556958436966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.13580834865570068, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.12273839861154556, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11844944953918457, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10661231726408005, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06316402554512024, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.058342963457107544, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.07202249020338058, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06590014696121216, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06419923901557922, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0540926456451416, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05188789963722229, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03701868653297424, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03203514590859413, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03085111454129219, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.03056243807077408, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01866956241428852, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.016841448843479156, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01653767004609108, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.015153011307120323, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01497772429138422, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.010514569468796253, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01155273150652647, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.010161700658500195, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.009117254987359047, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11844944953918457, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, "best_option": { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11844944953918457, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.29.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.21700073778629303, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2040451616048813, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.19943542778491974, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18074193596839905, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10315326601266861, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09837280958890915, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11512018740177155, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10643071681261063, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10411220043897629, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09204676002264023, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08780976384878159, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05881728604435921, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05110140144824982, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04965189844369888, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04930901154875755, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.029448529705405235, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.025899123400449753, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02556711994111538, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.0236845463514328, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.023471446707844734, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015752509236335754, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016186753287911415, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015275469981133938, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011358518153429031, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11512018740177155, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11512018740177155, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.29.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.22204113006591797, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.20880787074565887, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.20422783493995667, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18525876104831696, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10600678622722626, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10116534680128098, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11796829849481583, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10903660953044891, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10696632415056229, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0946364551782608, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0904959887266159, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0607856810092926, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.052812494337558746, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05144593492150307, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05111263692378998, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03049711510539055, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.027538392692804337, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02723059058189392, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025381695479154587, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02518070675432682, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01683741621673107, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.018144922330975533, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016393054276704788, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.013793304562568665, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11796829849481583, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11796829849481583, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.29.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.22857505083084106, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.20484140515327454, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.19437311589717865, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.17346076667308807, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.10481470823287964, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.09563370794057846, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.12606194615364075, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11560021340847015, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.10865890979766846, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09151481837034225, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08586464077234268, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.0642533153295517, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05557035282254219, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.050604648888111115, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.049383360892534256, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03244873881340027, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.026675600558519363, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.026343083009123802, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.024073094129562378, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02319105714559555, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.017646124586462975, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.017584413290023804, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.016032082960009575, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012335662730038166, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11560021340847015, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11560021340847015, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11160916090011597, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10457350313663483, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10140931606292725, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09147459268569946, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.052935317158699036, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04998664930462837, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06054263934493065, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.055878087878227234, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.053459420800209045, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04705487936735153, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04476465657353401, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.030794424936175346, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026757219806313515, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025437550619244576, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.025114042684435844, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015436827205121517, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01311676949262619, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012853975407779217, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01191430538892746, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011715607717633247, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.00811256654560566, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008071442134678364, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007678620517253876, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005326748359948397, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11160916090011597, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11160916090011597, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.10719342529773712, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.10002796351909637, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.09643544256687164, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.0868467390537262, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.05045505613088608, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04732671007514, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05839579179883003, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.054223258048295975, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.050958894193172455, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.04469594731926918, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.04263393208384514, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.029655905440449715, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.025917820632457733, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.024194158613681793, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.02380163036286831, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.014804276637732983, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.012330222874879837, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.011998724192380905, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.011138495989143848, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.01086329948157072, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.007677628193050623, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.00746657932177186, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.007142346352338791, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004576731938868761, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.10719342529773712, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.10719342529773712, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.27574995160102844, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.2581194043159485, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.25202685594558716, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.22708474099636078, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.13060158491134644, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.12415217608213425, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.1463432013988495, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.1345721185207367, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.1318860948085785, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11551252752542496, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.10984420776367188, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.07454828917980194, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.06429166346788406, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.06242042034864426, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.061988621950149536, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.0372125618159771, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.0315132662653923, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.03106936626136303, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.02835860848426819, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.02807691879570484, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.018999867141246796, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.018155844882130623, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.018188493326306343, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.010702386498451233, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11551252752542496, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.11551252752542496, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.17324577271938324, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.15598691999912262, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1500960737466812, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12764355540275574, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.080561563372612, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07464694231748581, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09158803522586823, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.08403663337230682, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08175810426473618, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0665525570511818, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06132371723651886, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04711362347006798, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.040724217891693115, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03916871175169945, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.03881928324699402, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.023807251825928688, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.021152924746274948, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02077382244169712, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.018520208075642586, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.0182893518358469, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.013303505256772041, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.014214039780199528, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.012844459153711796, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010954681783914566, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09158803522586823, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09158803522586823, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.20060467720031738, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18860718607902527, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18419668078422546, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.166909858584404, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09538966417312622, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09089236706495285, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10661400854587555, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09857963770627975, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09629058837890625, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0851045772433281, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0811614915728569, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.054509811103343964, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.047345347702503204, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.045906271785497665, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.0455622598528862, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.027259981259703636, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02391846850514412, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.023598749190568924, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02186015620827675, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02164294756948948, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014521600678563118, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014927585609257221, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.014034312218427658, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010417786426842213, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10661400854587555, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10661400854587555, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.186608225107193, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17542441189289093, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.17144142091274261, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15535451471805573, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0886433944106102, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08451981097459793, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09887801110744476, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09145016968250275, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08947063237428665, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07906326651573181, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07534073293209076, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05043259263038635, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0437726154923439, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.042511239647865295, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.0422145314514637, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.025179633870720863, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021875372156500816, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.021579565480351448, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019925585016608238, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.019734736531972885, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013182124122977257, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013233042322099209, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012750682421028614, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.008766246028244495, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09887801110744476, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09887801110744476, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.13113704323768616, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.11589188873767853, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.10847902297973633, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.09716574847698212, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.06006351485848427, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.05365963652729988, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.07506664097309113, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.06811325997114182, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.06267061829566956, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.05212818458676338, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.04939207807183266, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.03827255219221115, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.03300454840064049, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.029333721846342087, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.02841603197157383, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.019329050555825233, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.016002679243683815, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.015749268233776093, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.014530847780406475, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.013921777717769146, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.010647520422935486, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01130218431353569, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.009457978419959545, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.008579467423260212, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.11589188873767853, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, "best_option": { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.11589188873767853, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10337460041046143, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09680384397506714, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09392008185386658, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08487273007631302, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.049062952399253845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04636809602379799, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0562886968255043, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05161063373088837, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04955840855836868, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04356665536761284, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04170406609773636, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.02864142693579197, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02470877394080162, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.023551009595394135, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.023280801251530647, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01432883832603693, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01207633689045906, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011845848523080349, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010955280624330044, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01077328436076641, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007471572607755661, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007331605535000563, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007079911418259144, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.00471863616257906, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10337460041046143, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10337460041046143, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.k_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.10266613215208054, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.09584666043519974, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.0925372764468193, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.08326748013496399, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.04839322343468666, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.04541842266917229, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.05574708431959152, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.05165458098053932, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.04891504347324371, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.04277939721941948, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.04077141731977463, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.028250616043806076, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.024682117626070976, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.02320966124534607, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.022845979779958725, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.014127243310213089, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.011807464994490147, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.01152314618229866, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.010648450814187527, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.010416795499622822, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.007282603066414595, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.007113026455044746, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.006802669260650873, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.004361310042440891, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.10266613215208054, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.10266613215208054, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.v_proj", "numel": 2097152, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1904296875, "total_bits": 4593664.0, "err": 0.2474520206451416, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3779296875, "total_bits": 4986880.0, "err": 0.23109877109527588, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6279296875, "total_bits": 5511168.0, "err": 0.22523200511932373, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7216796875, "total_bits": 5707776.0, "err": 0.20287638902664185, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2216796875, "total_bits": 6756352.0, "err": 0.11703887581825256, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6904296875, "total_bits": 7739392.0, "err": 0.11099270731210709, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031982421875, "total_bits": 6358528.0, "err": 0.13174505531787872, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.12116539478302002, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11815040558576584, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5341796875, "total_bits": 7411712.0, "err": 0.10344216972589493, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65771484375, "total_bits": 7670784.0, "err": 0.0981357991695404, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031982421875, "total_bits": 8455680.0, "err": 0.06683717668056488, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1279296875, "total_bits": 8656896.0, "err": 0.057846345007419586, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2216796875, "total_bits": 8853504.0, "err": 0.05597122013568878, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3154296875, "total_bits": 9050112.0, "err": 0.05551084131002426, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031982421875, "total_bits": 10552832.0, "err": 0.033367402851581573, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2216796875, "total_bits": 10950656.0, "err": 0.02826811745762825, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3779296875, "total_bits": 11278336.0, "err": 0.027810849249362946, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5341796875, "total_bits": 11606016.0, "err": 0.025417078286409378, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7216796875, "total_bits": 11999232.0, "err": 0.02511082962155342, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031982421875, "total_bits": 12649984.0, "err": 0.017020802944898605, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1279296875, "total_bits": 12851200.0, "err": 0.016336148604750633, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.219482421875, "total_bits": 13043200.0, "err": 0.01627167873084545, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1279296875, "total_bits": 17045504.0, "err": 0.009655184112489223, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1279296875, "total_bits": 6559744.0, "err": 0.12116539478302002, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1904296875, "total_bits": 6690816.0, "err": 0.11815040558576584, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.059686996042728424, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.05246291682124138, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.04989033192396164, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.041081223636865616, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.027607396245002747, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.025114431977272034, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.03160784766077995, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.028882712125778198, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.028148574754595757, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.022041434422135353, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.01984722912311554, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.016310613602399826, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.014564480632543564, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.013988249003887177, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.01386008970439434, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.00859706662595272, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.008359150029718876, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.008222480304539204, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.00746142165735364, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.007382107898592949, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0053613511845469475, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.00645537069067359, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.005212740506976843, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005665201228111982, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.059686996042728424, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.059686996042728424, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.14665228128433228, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1378316432237625, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.13406041264533997, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.12152732908725739, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.07040634006261826, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.06680348515510559, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08039350807666779, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.07376209646463394, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07110419869422913, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.06296969205141068, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.060212887823581696, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04156733304262161, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03638962283730507, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03490638732910156, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.034552156925201416, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.021110540255904198, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.019729049876332283, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.019456692039966583, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.018405823037028313, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01820106990635395, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012188253924250603, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01427493616938591, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01170088816434145, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.011852031573653221, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.12152732908725739, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.12152732908725739, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.15723107755184174, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.147917240858078, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.14400677382946014, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.13057270646095276, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.07561307400465012, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07179809361696243, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08587971329689026, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.0788804143667221, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07627250999212265, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.06762567162513733, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.06467284262180328, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044338684529066086, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.039122574031353, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03764299675822258, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.03727605938911438, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.022686442360281944, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021599115803837776, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02130984514951706, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02020850032567978, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.019993776455521584, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013347454369068146, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015913859009742737, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012839389964938164, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01344187743961811, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08587971329689026, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08587971329689026, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.05507740378379822, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.0478755384683609, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.04183822125196457, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.037639617919921875, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.025530626997351646, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.02100612036883831, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.037263669073581696, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.032895639538764954, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.026992356404662132, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.022352147847414017, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.022072887048125267, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.019702419638633728, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.01664801687002182, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.013171084225177765, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.01224417332559824, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.010339174419641495, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.008187034167349339, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.007868747226893902, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.007694981060922146, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.007216591853648424, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.006204869598150253, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.00690238643437624, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.005142917390912771, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.00590673740953207, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.05507740378379822, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.05507740378379822, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } } ], "last_module_idx": 66, "base_perplexity": 8.352980873941503 }