diff --git "a/measurement-Yi-6B-Spicyboros-3.1.json" "b/measurement-Yi-6B-Spicyboros-3.1.json" new file mode 100644--- /dev/null +++ "b/measurement-Yi-6B-Spicyboros-3.1.json" @@ -0,0 +1,103668 @@ +{ + "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": 0.03611881285905838, + "qparams": { + "group_size": 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": 0.03588543459773064, + "qparams": { + "group_size": 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": 0.01836748979985714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.01826639287173748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.018201274797320366, + "qparams": { + "group_size": 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": 0.011868488043546677, + "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": 0.04828125238418579, + "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": 0.035584449768066406, + "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": 0.01821482740342617, + "qparams": { + "group_size": 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": 0.018159331753849983, + "qparams": { + "group_size": 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": 0.018018396571278572, + "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": 0.01931736432015896, + "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": 0.018141092732548714, + "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": 0.013258161023259163, + "qparams": { + "group_size": 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": 0.011831617914140224, + "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": 0.012536546215415001, + "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": 0.011821974068880081, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01133993361145258, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011819826439023018, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011338381096720695, + "qparams": { + "group_size": 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": 0.010939925909042358, + "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": 0.011819208040833473, + "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": 0.010401438921689987, + "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": 0.011337260715663433, + "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": 0.03611881285905838, + "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": 0.03611881285905838, + "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": 0.03649970516562462, + "qparams": { + "group_size": 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": 0.03626677021384239, + "qparams": { + "group_size": 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": 0.015028552152216434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.014938484877347946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.014881428331136703, + "qparams": { + "group_size": 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": 0.008572345599532127, + "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": 0.04723719134926796, + "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": 0.035961952060461044, + "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": 0.014892831444740295, + "qparams": { + "group_size": 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": 0.014840621501207352, + "qparams": { + "group_size": 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": 0.014972376637160778, + "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": 0.018379172310233116, + "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": 0.014826537109911442, + "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": 0.00982783455401659, + "qparams": { + "group_size": 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": 0.00853940099477768, + "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": 0.009729765355587006, + "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": 0.008529414422810078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008199188858270645, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008525007404386997, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008197230286896229, + "qparams": { + "group_size": 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": 0.008246841840445995, + "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": 0.008524748496711254, + "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": 0.007791536394506693, + "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": 0.008195287548005581, + "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": 0.03649970516562462, + "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": 0.03649970516562462, + "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": 0.1323116570711136, + "qparams": { + "group_size": 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": 0.1130238026380539, + "qparams": { + "group_size": 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": 0.09275054931640625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07566187530755997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06125125661492348, + "qparams": { + "group_size": 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": 0.04615986347198486, + "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": 0.11703991144895554, + "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": 0.08365976810455322, + "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": 0.06318357586860657, + "qparams": { + "group_size": 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": 0.04860250651836395, + "qparams": { + "group_size": 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": 0.05369417369365692, + "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": 0.06746789813041687, + "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": 0.040399219840765, + "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": 0.029611684381961823, + "qparams": { + "group_size": 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": 0.026489177718758583, + "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": 0.035426653921604156, + "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": 0.015416368842124939, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.013384808786213398, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.012923029251396656, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010723523795604706, + "qparams": { + "group_size": 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": 0.018267309293150902, + "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": 0.011365694925189018, + "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": 0.008625169284641743, + "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": 0.006731962785124779, + "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": 0.09275054931640625, + "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": 0.09275054931640625, + "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": 0.016372067853808403, + "qparams": { + "group_size": 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": 0.014088345691561699, + "qparams": { + "group_size": 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": 0.01316052209585905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.010780286975204945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.007341309450566769, + "qparams": { + "group_size": 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": 0.006468828767538071, + "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": 0.009214297868311405, + "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": 0.008132703602313995, + "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": 0.007479531690478325, + "qparams": { + "group_size": 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": 0.005885310471057892, + "qparams": { + "group_size": 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": 0.0053910729475319386, + "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": 0.004670414607971907, + "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": 0.0041355108842253685, + "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": 0.00379947223700583, + "qparams": { + "group_size": 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": 0.0037136629689484835, + "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": 0.002559620188549161, + "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": 0.0024516547564417124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.002391391433775425, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0022477353923022747, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0022065744269639254, + "qparams": { + "group_size": 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": 0.001717306557111442, + "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": 0.0020471864845603704, + "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": 0.0016331124352291226, + "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": 0.0018551526591181755, + "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": 0.016372067853808403, + "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": 0.016372067853808403, + "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": 0.08406222611665726, + "qparams": { + "group_size": 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": 0.08105499297380447, + "qparams": { + "group_size": 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": 0.08010488003492355, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07481254637241364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.038159437477588654, + "qparams": { + "group_size": 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": 0.0373513288795948, + "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": 0.04177072271704674, + "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": 0.03869757056236267, + "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": 0.03832276165485382, + "qparams": { + "group_size": 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": 0.035937096923589706, + "qparams": { + "group_size": 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": 0.03525794297456741, + "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": 0.021096518263220787, + "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": 0.01840590499341488, + "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": 0.018186617642641068, + "qparams": { + "group_size": 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": 0.01813320443034172, + "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": 0.01053438801318407, + "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": 0.009509982541203499, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009461671113967896, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.009093147702515125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009066053666174412, + "qparams": { + "group_size": 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": 0.0056484537199139595, + "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": 0.0059454794973134995, + "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": 0.005581056699156761, + "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": 0.004282958339899778, + "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": 2.1876362645348837, + "total_bits": 98637824.0, + "err": 0.08406222611665726, + "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": 2.1876362645348837, + "total_bits": 98637824.0, + "err": 0.08406222611665726, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.13042861223220825, + "qparams": { + "group_size": 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": 0.12604857981204987, + "qparams": { + "group_size": 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": 0.12470489740371704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.116752490401268, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05951666086912155, + "qparams": { + "group_size": 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": 0.05833379179239273, + "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": 0.06504950672388077, + "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": 0.06021316349506378, + "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": 0.059753164649009705, + "qparams": { + "group_size": 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": 0.056113190948963165, + "qparams": { + "group_size": 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": 0.05501333251595497, + "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": 0.032770801335573196, + "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": 0.028491146862506866, + "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": 0.028198644518852234, + "qparams": { + "group_size": 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": 0.02813086472451687, + "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": 0.016318311914801598, + "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": 0.014328579418361187, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01425202377140522, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.013659052550792694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.013617769815027714, + "qparams": { + "group_size": 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": 0.008466754108667374, + "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": 0.008329964242875576, + "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": 0.008365507237613201, + "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": 0.0052596451714634895, + "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": 0.06504950672388077, + "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": 0.06504950672388077, + "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": 0.06334716826677322, + "qparams": { + "group_size": 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": 0.058086201548576355, + "qparams": { + "group_size": 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": 0.05566412955522537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05088726058602333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.028298448771238327, + "qparams": { + "group_size": 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": 0.02632608637213707, + "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": 0.03364015743136406, + "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": 0.0309695266187191, + "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": 0.029088908806443214, + "qparams": { + "group_size": 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": 0.025642510503530502, + "qparams": { + "group_size": 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": 0.02460477501153946, + "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": 0.017096318304538727, + "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": 0.014880440197885036, + "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": 0.013679146766662598, + "qparams": { + "group_size": 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": 0.01338876411318779, + "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": 0.008605961687862873, + "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": 0.007426551077514887, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00733597669750452, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006940115708857775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006745882797986269, + "qparams": { + "group_size": 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": 0.004746180959045887, + "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": 0.005138975568115711, + "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": 0.004369653295725584, + "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": 0.003928951919078827, + "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": 0.06334716826677322, + "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": 0.06334716826677322, + "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": 0.010809332132339478, + "qparams": { + "group_size": 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": 0.010612276382744312, + "qparams": { + "group_size": 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": 0.005431319586932659, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.005364769138395786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.005323946941643953, + "qparams": { + "group_size": 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": 0.0027666143141686916, + "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": 0.011945312842726707, + "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": 0.010494858957827091, + "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": 0.005343427881598473, + "qparams": { + "group_size": 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": 0.0052624912932515144, + "qparams": { + "group_size": 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": 0.005476873833686113, + "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": 0.005768630653619766, + "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": 0.005247816909104586, + "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": 0.003336975583806634, + "qparams": { + "group_size": 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": 0.00270414468832314, + "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": 0.0033113642130047083, + "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": 0.002675711177289486, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0024604727514088154, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0026704620104283094, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0024553188122808933, + "qparams": { + "group_size": 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": 0.0025134875904768705, + "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": 0.0026692142710089684, + "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": 0.002230520825833082, + "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": 0.002451632171869278, + "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": 0.010809332132339478, + "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": 0.010809332132339478, + "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": 0.008402074687182903, + "qparams": { + "group_size": 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": 0.008224410936236382, + "qparams": { + "group_size": 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": 0.004062020219862461, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.003986596595495939, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.003944705706089735, + "qparams": { + "group_size": 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": 0.0017587692709639668, + "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": 0.009271453134715557, + "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": 0.008110616356134415, + "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": 0.0039621018804609776, + "qparams": { + "group_size": 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": 0.003884645877406001, + "qparams": { + "group_size": 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": 0.004049889277666807, + "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": 0.004313095938414335, + "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": 0.0038690820802003145, + "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": 0.0022762762382626534, + "qparams": { + "group_size": 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": 0.0016818225849419832, + "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": 0.0023195466492325068, + "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": 0.0016457210294902325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0014320001937448978, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0016385846538469195, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0014247539220377803, + "qparams": { + "group_size": 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": 0.0016132405726239085, + "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": 0.0016368379583582282, + "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": 0.0013505752431228757, + "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": 0.0014198091812431812, + "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": 0.008402074687182903, + "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": 0.008402074687182903, + "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": 0.1158689484000206, + "qparams": { + "group_size": 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": 0.09632052481174469, + "qparams": { + "group_size": 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": 0.08146616816520691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06640756875276566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.052654191851615906, + "qparams": { + "group_size": 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": 0.040693499147892, + "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": 0.09929832816123962, + "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": 0.06807174533605576, + "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": 0.05440039187669754, + "qparams": { + "group_size": 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": 0.04076346009969711, + "qparams": { + "group_size": 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": 0.04524675011634827, + "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": 0.055538423359394073, + "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": 0.03289983049035072, + "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": 0.025341076776385307, + "qparams": { + "group_size": 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": 0.023272503167390823, + "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": 0.028778014704585075, + "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": 0.013241082429885864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0118241673335433, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.010897394269704819, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009427734650671482, + "qparams": { + "group_size": 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": 0.014506803825497627, + "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": 0.009472268633544445, + "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": 0.007496200501918793, + "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": 0.005815728567540646, + "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": 0.09632052481174469, + "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": 0.09632052481174469, + "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": 0.04038184881210327, + "qparams": { + "group_size": 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": 0.03207200765609741, + "qparams": { + "group_size": 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": 0.029064493253827095, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.023217711597681046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.017517654225230217, + "qparams": { + "group_size": 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": 0.01466510072350502, + "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": 0.022208545356988907, + "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": 0.019671203568577766, + "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": 0.018213018774986267, + "qparams": { + "group_size": 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": 0.012921546585857868, + "qparams": { + "group_size": 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": 0.012291896156966686, + "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": 0.011379752308130264, + "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": 0.010074243880808353, + "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": 0.00912262499332428, + "qparams": { + "group_size": 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": 0.00889698974788189, + "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": 0.006238183937966824, + "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": 0.006008273456245661, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.005865270271897316, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.005420461762696505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.005304887890815735, + "qparams": { + "group_size": 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": 0.004232750274240971, + "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": 0.005112974904477596, + "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": 0.004010261036455631, + "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": 0.004682155326008797, + "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": 0.04038184881210327, + "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": 0.04038184881210327, + "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": 0.0947185605764389, + "qparams": { + "group_size": 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": 0.09093393385410309, + "qparams": { + "group_size": 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": 0.08968476951122284, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08342427760362625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.043136660009622574, + "qparams": { + "group_size": 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": 0.04203227534890175, + "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": 0.047305308282375336, + "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": 0.04387175664305687, + "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": 0.043362077325582504, + "qparams": { + "group_size": 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": 0.04029951989650726, + "qparams": { + "group_size": 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": 0.03939816728234291, + "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": 0.0240451842546463, + "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": 0.021079285070300102, + "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": 0.020783448591828346, + "qparams": { + "group_size": 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": 0.02071332186460495, + "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": 0.012047876603901386, + "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": 0.011187916621565819, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011123007163405418, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01067259069532156, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010632941499352455, + "qparams": { + "group_size": 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": 0.006669366266578436, + "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": 0.007435276638716459, + "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": 0.006578623317182064, + "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": 0.0057757580652832985, + "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": 2.1876362645348837, + "total_bits": 98637824.0, + "err": 0.0947185605764389, + "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": 2.1876362645348837, + "total_bits": 98637824.0, + "err": 0.0947185605764389, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.11623897403478622, + "qparams": { + "group_size": 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": 0.1117359921336174, + "qparams": { + "group_size": 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": 0.11027596145868301, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10274133831262589, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.052926935255527496, + "qparams": { + "group_size": 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": 0.05161377787590027, + "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": 0.057973433285951614, + "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": 0.05371362715959549, + "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": 0.05318571627140045, + "qparams": { + "group_size": 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": 0.04947785660624504, + "qparams": { + "group_size": 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": 0.048312265425920486, + "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": 0.029227616265416145, + "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": 0.02544354647397995, + "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": 0.025109779089689255, + "qparams": { + "group_size": 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": 0.025024520233273506, + "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": 0.01457409467548132, + "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": 0.01280998345464468, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01272191945463419, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.012131545692682266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012080077081918716, + "qparams": { + "group_size": 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": 0.0076236724853515625, + "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": 0.007549365982413292, + "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": 0.007505355402827263, + "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": 0.0048740035854279995, + "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": 0.057973433285951614, + "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": 0.057973433285951614, + "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": 0.04995548725128174, + "qparams": { + "group_size": 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": 0.04341917112469673, + "qparams": { + "group_size": 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": 0.03027857467532158, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.02768130972981453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.021138109266757965, + "qparams": { + "group_size": 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": 0.0148507384583354, + "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": 0.03568752482533455, + "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": 0.0362006276845932, + "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": 0.02249211259186268, + "qparams": { + "group_size": 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": 0.018723761662840843, + "qparams": { + "group_size": 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": 0.017606966197490692, + "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": 0.01658768020570278, + "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": 0.015991806983947754, + "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": 0.010754307731986046, + "qparams": { + "group_size": 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": 0.009549476206302643, + "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": 0.010566932149231434, + "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": 0.007038225419819355, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.005387107841670513, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.00667900824919343, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.004754464607685804, + "qparams": { + "group_size": 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": 0.006315390579402447, + "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": 0.006305975373834372, + "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": 0.003877912648022175, + "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": 0.0038956552743911743, + "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": 0.04995548725128174, + "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": 0.04995548725128174, + "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": 0.02415706403553486, + "qparams": { + "group_size": 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": 0.02101081982254982, + "qparams": { + "group_size": 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": 0.016376778483390808, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.014638560824096203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.011009004898369312, + "qparams": { + "group_size": 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": 0.007833611220121384, + "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": 0.02153823897242546, + "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": 0.016302665695548058, + "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": 0.011406514793634415, + "qparams": { + "group_size": 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": 0.009456473402678967, + "qparams": { + "group_size": 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": 0.009970493614673615, + "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": 0.011296454817056656, + "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": 0.007947376929223537, + "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": 0.005477269180119038, + "qparams": { + "group_size": 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": 0.004722141660749912, + "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": 0.005671774037182331, + "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": 0.003125788876786828, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0027018480468541384, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.002858659951016307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.002422984689474106, + "qparams": { + "group_size": 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": 0.0029163432773202658, + "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": 0.002616677898913622, + "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": 0.0016255740774795413, + "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": 0.0019156734924763441, + "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": 0.02415706403553486, + "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": 0.02415706403553486, + "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": 0.02315150573849678, + "qparams": { + "group_size": 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": 0.020031370222568512, + "qparams": { + "group_size": 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": 0.014902393333613873, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.013391311280429363, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.010352171957492828, + "qparams": { + "group_size": 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": 0.006950560957193375, + "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": 0.017711758613586426, + "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": 0.016068946570158005, + "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": 0.010769259184598923, + "qparams": { + "group_size": 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": 0.008929699659347534, + "qparams": { + "group_size": 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": 0.0092544574290514, + "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": 0.009034368209540844, + "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": 0.007726586423814297, + "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": 0.005071511026471853, + "qparams": { + "group_size": 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": 0.004244414158165455, + "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": 0.0045462665148079395, + "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": 0.0027890782803297043, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.002307323506101966, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.002528796438127756, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.002022089436650276, + "qparams": { + "group_size": 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": 0.002387135988101363, + "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": 0.0023193301167339087, + "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": 0.0014998827828094363, + "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": 0.0015356764197349548, + "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": 0.02315150573849678, + "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": 0.02315150573849678, + "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": 0.14455506205558777, + "qparams": { + "group_size": 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": 0.12850777804851532, + "qparams": { + "group_size": 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": 0.12141916155815125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10701191425323486, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06691648066043854, + "qparams": { + "group_size": 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": 0.059991199523210526, + "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": 0.07989978790283203, + "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": 0.07342839241027832, + "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": 0.06835047900676727, + "qparams": { + "group_size": 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": 0.05606679245829582, + "qparams": { + "group_size": 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": 0.053000207990407944, + "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": 0.04089701920747757, + "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": 0.03506329655647278, + "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": 0.03208347037434578, + "qparams": { + "group_size": 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": 0.031334519386291504, + "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": 0.020441381260752678, + "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": 0.016317525878548622, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01568821631371975, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.014080578461289406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.013554812408983707, + "qparams": { + "group_size": 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": 0.010425961576402187, + "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": 0.010007045231759548, + "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": 0.009246156550943851, + "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": 0.005979449022561312, + "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": 0.07989978790283203, + "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": 0.07989978790283203, + "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": 0.18458110094070435, + "qparams": { + "group_size": 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": 0.1388942301273346, + "qparams": { + "group_size": 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": 0.12145569920539856, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10804261267185211, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08345917612314224, + "qparams": { + "group_size": 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": 0.06448774039745331, + "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": 0.10320591926574707, + "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": 0.09384337067604065, + "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": 0.08726463466882706, + "qparams": { + "group_size": 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": 0.058925457298755646, + "qparams": { + "group_size": 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": 0.0567069947719574, + "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": 0.052937231957912445, + "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": 0.046265654265880585, + "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": 0.04171154648065567, + "qparams": { + "group_size": 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": 0.040602900087833405, + "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": 0.027330465614795685, + "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": 0.02406969666481018, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023024901747703552, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01999749429523945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019307933747768402, + "qparams": { + "group_size": 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": 0.016235485672950745, + "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": 0.018217429518699646, + "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": 0.014960973523557186, + "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": 0.015185198746621609, + "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": 0.09384337067604065, + "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": 0.09384337067604065, + "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": 0.15065859258174896, + "qparams": { + "group_size": 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": 0.14405792951583862, + "qparams": { + "group_size": 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": 0.14206908643245697, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.13141632080078125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06876951456069946, + "qparams": { + "group_size": 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": 0.06685025244951248, + "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": 0.07575605809688568, + "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": 0.06986823678016663, + "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": 0.06915470957756042, + "qparams": { + "group_size": 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": 0.06377354264259338, + "qparams": { + "group_size": 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": 0.062115058302879333, + "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": 0.03839889168739319, + "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": 0.03329770639538765, + "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": 0.03280135989189148, + "qparams": { + "group_size": 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": 0.03268902003765106, + "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": 0.01920202374458313, + "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": 0.01713535375893116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01701367273926735, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.016175948083400726, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.016113298013806343, + "qparams": { + "group_size": 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": 0.010381743311882019, + "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": 0.010693398304283619, + "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": 0.010223684832453728, + "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": 0.007646495476365089, + "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": 0.07575605809688568, + "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": 0.07575605809688568, + "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": 0.17664030194282532, + "qparams": { + "group_size": 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": 0.1693175584077835, + "qparams": { + "group_size": 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": 0.16701489686965942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15484225749969482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08102540671825409, + "qparams": { + "group_size": 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": 0.07883897423744202, + "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": 0.0890379250049591, + "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": 0.08230810612440109, + "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": 0.08147671073675156, + "qparams": { + "group_size": 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": 0.07523710280656815, + "qparams": { + "group_size": 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": 0.0732233077287674, + "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": 0.04496089741587639, + "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": 0.0390237532556057, + "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": 0.0384712815284729, + "qparams": { + "group_size": 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": 0.03833010047674179, + "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": 0.02241677977144718, + "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": 0.019559545442461967, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01941828802227974, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01840117946267128, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018313460052013397, + "qparams": { + "group_size": 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": 0.011669116094708443, + "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": 0.011414634995162487, + "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": 0.011485266499221325, + "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": 0.00720244599506259, + "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": 0.0890379250049591, + "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": 0.0890379250049591, + "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": 0.15317212045192719, + "qparams": { + "group_size": 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": 0.13932910561561584, + "qparams": { + "group_size": 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": 0.1339561939239502, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12201951444149017, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06947468221187592, + "qparams": { + "group_size": 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": 0.06471790373325348, + "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": 0.08177056163549423, + "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": 0.07452868670225143, + "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": 0.07151399552822113, + "qparams": { + "group_size": 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": 0.06217906251549721, + "qparams": { + "group_size": 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": 0.05934092029929161, + "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": 0.04140975698828697, + "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": 0.03578270599246025, + "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": 0.03352753818035126, + "qparams": { + "group_size": 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": 0.03297116979956627, + "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": 0.020999273285269737, + "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": 0.017861057072877884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01774320751428604, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.016472352668642998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01607659086585045, + "qparams": { + "group_size": 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": 0.011696147732436657, + "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": 0.011865022592246532, + "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": 0.010992845520377159, + "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": 0.008737538941204548, + "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": 0.08177056163549423, + "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": 0.08177056163549423, + "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": 0.03212513402104378, + "qparams": { + "group_size": 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": 0.02887757495045662, + "qparams": { + "group_size": 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": 0.0256455410271883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.02274494245648384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.015008989721536636, + "qparams": { + "group_size": 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": 0.01262317318469286, + "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": 0.02640562504529953, + "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": 0.0188748799264431, + "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": 0.015315424650907516, + "qparams": { + "group_size": 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": 0.012847189791500568, + "qparams": { + "group_size": 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": 0.01276999432593584, + "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": 0.013727151788771152, + "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": 0.00914614275097847, + "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": 0.007334096822887659, + "qparams": { + "group_size": 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": 0.006836950313299894, + "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": 0.006857687141746283, + "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": 0.003958724904805422, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0036491602659225464, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.00355037534609437, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0032424291130155325, + "qparams": { + "group_size": 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": 0.003553261747583747, + "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": 0.0029067625291645527, + "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": 0.0022919857874512672, + "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": 0.002047584392130375, + "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": 0.03212513402104378, + "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": 0.03212513402104378, + "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": 0.03167732059955597, + "qparams": { + "group_size": 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": 0.028106244280934334, + "qparams": { + "group_size": 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": 0.024445723742246628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.021571850404143333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.014546814374625683, + "qparams": { + "group_size": 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": 0.011930732056498528, + "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": 0.020903922617435455, + "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": 0.01886020042002201, + "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": 0.014887146651744843, + "qparams": { + "group_size": 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": 0.012304519303143024, + "qparams": { + "group_size": 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": 0.01207140926271677, + "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": 0.010556657798588276, + "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": 0.009000992402434349, + "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": 0.0070691099390387535, + "qparams": { + "group_size": 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": 0.0065171364694833755, + "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": 0.00528738321736455, + "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": 0.0037465107161551714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.003400162560865283, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.003318848554044962, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0029668896459043026, + "qparams": { + "group_size": 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": 0.0027588873635977507, + "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": 0.0027210514526814222, + "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": 0.0020587993785738945, + "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": 0.0017948502209037542, + "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": 0.03167732059955597, + "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": 0.03167732059955597, + "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": 0.16588501632213593, + "qparams": { + "group_size": 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": 0.15149414539337158, + "qparams": { + "group_size": 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": 0.1456012725830078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1294628381729126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07777643203735352, + "qparams": { + "group_size": 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": 0.07199627161026001, + "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": 0.09129434078931808, + "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": 0.08282111585140228, + "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": 0.07895829528570175, + "qparams": { + "group_size": 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": 0.06699660420417786, + "qparams": { + "group_size": 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": 0.06361076235771179, + "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": 0.04673616588115692, + "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": 0.03964484855532646, + "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": 0.037282370030879974, + "qparams": { + "group_size": 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": 0.036712855100631714, + "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": 0.023361586034297943, + "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": 0.018874753266572952, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018373433500528336, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.016605503857135773, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01622183993458748, + "qparams": { + "group_size": 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": 0.011915815062820911, + "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": 0.011202494613826275, + "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": 0.010798431001603603, + "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": 0.00664867227897048, + "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": 0.09129434078931808, + "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": 0.09129434078931808, + "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": 0.15233123302459717, + "qparams": { + "group_size": 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": 0.1238650530576706, + "qparams": { + "group_size": 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": 0.11292128264904022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09176076203584671, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06989389657974243, + "qparams": { + "group_size": 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": 0.05967167392373085, + "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": 0.08532094210386276, + "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": 0.07717645913362503, + "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": 0.07209089398384094, + "qparams": { + "group_size": 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": 0.05130063742399216, + "qparams": { + "group_size": 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": 0.047195713967084885, + "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": 0.04399387910962105, + "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": 0.03815162926912308, + "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": 0.034781720489263535, + "qparams": { + "group_size": 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": 0.03398742154240608, + "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": 0.022574525326490402, + "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": 0.019962621852755547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01933397725224495, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01685168966650963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01634758710861206, + "qparams": { + "group_size": 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": 0.01308727078139782, + "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": 0.014964070171117783, + "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": 0.012027339078485966, + "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": 0.012442675419151783, + "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": 0.09176076203584671, + "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": 0.09176076203584671, + "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": 0.1759834587574005, + "qparams": { + "group_size": 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": 0.16748104989528656, + "qparams": { + "group_size": 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": 0.16480273008346558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15147507190704346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08204956352710724, + "qparams": { + "group_size": 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": 0.07924710214138031, + "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": 0.09038553386926651, + "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": 0.08369871228933334, + "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": 0.08262398838996887, + "qparams": { + "group_size": 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": 0.07502469420433044, + "qparams": { + "group_size": 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": 0.0723884254693985, + "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": 0.045917246490716934, + "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": 0.03996706008911133, + "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": 0.03920217230916023, + "qparams": { + "group_size": 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": 0.03902720659971237, + "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": 0.022939983755350113, + "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": 0.020229000598192215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020043620839715004, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01882878690958023, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018717437982559204, + "qparams": { + "group_size": 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": 0.012092563323676586, + "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": 0.012275533750653267, + "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": 0.01183678675442934, + "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": 0.00831962563097477, + "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": 0.09038553386926651, + "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": 0.09038553386926651, + "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": 0.20259793102741241, + "qparams": { + "group_size": 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": 0.19290348887443542, + "qparams": { + "group_size": 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": 0.18975567817687988, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1745244264602661, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09436554461717606, + "qparams": { + "group_size": 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": 0.09117982536554337, + "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": 0.10392089933156967, + "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": 0.09623752534389496, + "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": 0.09503012895584106, + "qparams": { + "group_size": 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": 0.08633444458246231, + "qparams": { + "group_size": 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": 0.0832521840929985, + "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": 0.052631329745054245, + "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": 0.04580061510205269, + "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": 0.0449453741312027, + "qparams": { + "group_size": 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": 0.04473811760544777, + "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": 0.026235191151499748, + "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": 0.022785069420933723, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022566305473446846, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021129228174686432, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020999955013394356, + "qparams": { + "group_size": 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": 0.013494192622601986, + "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": 0.013203758746385574, + "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": 0.013195618987083435, + "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": 0.008121546357870102, + "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": 0.09623752534389496, + "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": 0.09623752534389496, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "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": 0.10046497732400894, + "qparams": { + "group_size": 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": 0.09010004997253418, + "qparams": { + "group_size": 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": 0.08331573754549026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07513169944286346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04590771347284317, + "qparams": { + "group_size": 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": 0.040888819843530655, + "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": 0.05938588082790375, + "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": 0.05394991487264633, + "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": 0.04741957411170006, + "qparams": { + "group_size": 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": 0.04037823528051376, + "qparams": { + "group_size": 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": 0.03833690285682678, + "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": 0.029168197885155678, + "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": 0.025762563571333885, + "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": 0.022569239139556885, + "qparams": { + "group_size": 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": 0.021466128528118134, + "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": 0.01512968074530363, + "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": 0.01207280158996582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01161008607596159, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011034428142011166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010326258838176727, + "qparams": { + "group_size": 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": 0.008163325488567352, + "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": 0.008513832464814186, + "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": 0.007036999799311161, + "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": 0.006020219996571541, + "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": 0.10046497732400894, + "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": 0.10046497732400894, + "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": 0.056068629026412964, + "qparams": { + "group_size": 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": 0.05147848650813103, + "qparams": { + "group_size": 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": 0.04898291826248169, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04378329589962959, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.026334574446082115, + "qparams": { + "group_size": 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": 0.024175751954317093, + "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": 0.03178529813885689, + "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": 0.028921721503138542, + "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": 0.026730136945843697, + "qparams": { + "group_size": 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": 0.022875964641571045, + "qparams": { + "group_size": 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": 0.02200894057750702, + "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": 0.016258442774415016, + "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": 0.01386922225356102, + "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": 0.012687355279922485, + "qparams": { + "group_size": 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": 0.012402331456542015, + "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": 0.008139194920659065, + "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": 0.006609148811548948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006392231676727533, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0059145535342395306, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0057282657362520695, + "qparams": { + "group_size": 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": 0.004255462437868118, + "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": 0.0042467666789889336, + "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": 0.003790304996073246, + "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": 0.0028607004787772894, + "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": 0.056068629026412964, + "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": 0.056068629026412964, + "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": 0.061453673988580704, + "qparams": { + "group_size": 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": 0.055644553154706955, + "qparams": { + "group_size": 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": 0.052420634776353836, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04645490646362305, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.028481025248765945, + "qparams": { + "group_size": 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": 0.025759659707546234, + "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": 0.034508537501096725, + "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": 0.03175026550889015, + "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": 0.028963344171643257, + "qparams": { + "group_size": 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": 0.024324962869286537, + "qparams": { + "group_size": 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": 0.023267172276973724, + "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": 0.017498236149549484, + "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": 0.015147775411605835, + "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": 0.01367171946913004, + "qparams": { + "group_size": 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": 0.013294236734509468, + "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": 0.008760218508541584, + "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": 0.007000487297773361, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006717240437865257, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006157142575830221, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.005909544415771961, + "qparams": { + "group_size": 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": 0.004526685457676649, + "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": 0.0043909866362810135, + "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": 0.0040029981173574924, + "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": 0.002711919602006674, + "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": 0.061453673988580704, + "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": 0.061453673988580704, + "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": 0.1759043037891388, + "qparams": { + "group_size": 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": 0.16320399940013885, + "qparams": { + "group_size": 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": 0.15805374085903168, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14185385406017303, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08271987736225128, + "qparams": { + "group_size": 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": 0.07769815623760223, + "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": 0.09456716477870941, + "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": 0.08697547018527985, + "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": 0.08378466963768005, + "qparams": { + "group_size": 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": 0.07258288562297821, + "qparams": { + "group_size": 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": 0.06909497827291489, + "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": 0.04822688549757004, + "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": 0.04163774102926254, + "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": 0.03965150937438011, + "qparams": { + "group_size": 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": 0.03918967396020889, + "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": 0.02407943084836006, + "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": 0.020084206014871597, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019669756293296814, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.017964737489819527, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01764347031712532, + "qparams": { + "group_size": 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": 0.012321599759161472, + "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": 0.011852577328681946, + "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": 0.011542015708982944, + "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": 0.00710570951923728, + "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": 0.09456716477870941, + "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": 0.09456716477870941, + "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": 0.148682102560997, + "qparams": { + "group_size": 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": 0.12935848534107208, + "qparams": { + "group_size": 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": 0.12142068892717361, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0981510579586029, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0692400187253952, + "qparams": { + "group_size": 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": 0.061427611857652664, + "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": 0.08422379195690155, + "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": 0.07493594288825989, + "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": 0.07085330784320831, + "qparams": { + "group_size": 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": 0.05430126562714577, + "qparams": { + "group_size": 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": 0.04931120574474335, + "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": 0.04351295158267021, + "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": 0.0370083786547184, + "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": 0.03445731848478317, + "qparams": { + "group_size": 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": 0.033814311027526855, + "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": 0.02252073585987091, + "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": 0.01949288696050644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0189692210406065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.016883760690689087, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0165056474506855, + "qparams": { + "group_size": 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": 0.013450954109430313, + "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": 0.014282199554145336, + "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": 0.012639656662940979, + "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": 0.011764474213123322, + "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": 0.0981510579586029, + "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": 0.0981510579586029, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "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": 0.1881938874721527, + "qparams": { + "group_size": 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": 0.17862750589847565, + "qparams": { + "group_size": 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": 0.17561200261116028, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16092485189437866, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08816810697317123, + "qparams": { + "group_size": 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": 0.08499961346387863, + "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": 0.09732871502637863, + "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": 0.08998546004295349, + "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": 0.08882856369018555, + "qparams": { + "group_size": 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": 0.08019210398197174, + "qparams": { + "group_size": 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": 0.07722610235214233, + "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": 0.04952157288789749, + "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": 0.043010760098695755, + "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": 0.04219968989491463, + "qparams": { + "group_size": 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": 0.04200804606080055, + "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": 0.02481786347925663, + "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": 0.02181835286319256, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02161393128335476, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020218228921294212, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020098552107810974, + "qparams": { + "group_size": 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": 0.013200021348893642, + "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": 0.013279775157570839, + "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": 0.012924262322485447, + "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": 0.009034636430442333, + "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": 0.09732871502637863, + "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": 0.09732871502637863, + "qparams": { + "group_size": 128, + "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": 0.2206503003835678, + "qparams": { + "group_size": 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": 0.20958247780799866, + "qparams": { + "group_size": 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": 0.2060691863298416, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18889719247817993, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10333171486854553, + "qparams": { + "group_size": 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": 0.09963100403547287, + "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": 0.11385195702314377, + "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": 0.10541156679391861, + "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": 0.10407409816980362, + "qparams": { + "group_size": 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": 0.09397922456264496, + "qparams": { + "group_size": 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": 0.0903867706656456, + "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": 0.05770425871014595, + "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": 0.05021195486187935, + "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": 0.04927363991737366, + "qparams": { + "group_size": 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": 0.04905073344707489, + "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": 0.02876913733780384, + "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": 0.025010868906974792, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02476702816784382, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.023081177845597267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022938231006264687, + "qparams": { + "group_size": 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": 0.014848717488348484, + "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": 0.01452589314430952, + "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": 0.01453459169715643, + "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": 0.008980613201856613, + "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": 0.09397922456264496, + "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": 0.09397922456264496, + "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": 0.20303934812545776, + "qparams": { + "group_size": 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": 0.18429172039031982, + "qparams": { + "group_size": 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": 0.17697852849960327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16026078164577484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09272946417331696, + "qparams": { + "group_size": 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": 0.08613542467355728, + "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": 0.10830741375684738, + "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": 0.099665068089962, + "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": 0.09553761035203934, + "qparams": { + "group_size": 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": 0.08234471827745438, + "qparams": { + "group_size": 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": 0.07796633243560791, + "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": 0.05483630299568176, + "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": 0.04777808487415314, + "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": 0.044672127813100815, + "qparams": { + "group_size": 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": 0.0439143069088459, + "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": 0.02748611383140087, + "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": 0.023558277636766434, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023398814722895622, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021553391590714455, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02099921554327011, + "qparams": { + "group_size": 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": 0.014725454151630402, + "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": 0.01535802148282528, + "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": 0.013733143918216228, + "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": 0.01096798200160265, + "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": 0.099665068089962, + "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": 0.099665068089962, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "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": 0.0597061812877655, + "qparams": { + "group_size": 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": 0.05543745309114456, + "qparams": { + "group_size": 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": 0.05249997600913048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.047246433794498444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.028181983157992363, + "qparams": { + "group_size": 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": 0.025866176933050156, + "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": 0.03406975045800209, + "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": 0.0315374881029129, + "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": 0.02853856049478054, + "qparams": { + "group_size": 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": 0.024832772091031075, + "qparams": { + "group_size": 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": 0.023813461884856224, + "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": 0.01740201935172081, + "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": 0.015169798396527767, + "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": 0.013625104911625385, + "qparams": { + "group_size": 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": 0.013242670334875584, + "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": 0.008724628016352654, + "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": 0.0071777477860450745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006912172306329012, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006519002839922905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006280867848545313, + "qparams": { + "group_size": 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": 0.004627720918506384, + "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": 0.004765660502016544, + "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": 0.004129287786781788, + "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": 0.00331907463259995, + "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": 0.0597061812877655, + "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": 0.0597061812877655, + "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": 0.05087171494960785, + "qparams": { + "group_size": 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": 0.0469287671148777, + "qparams": { + "group_size": 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": 0.04425523430109024, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03959110379219055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.023721622303128242, + "qparams": { + "group_size": 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": 0.021635213866829872, + "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": 0.02914269268512726, + "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": 0.026903007179498672, + "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": 0.0240685623139143, + "qparams": { + "group_size": 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": 0.020815880969166756, + "qparams": { + "group_size": 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": 0.019887378439307213, + "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": 0.014734894968569279, + "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": 0.012865005061030388, + "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": 0.011435326188802719, + "qparams": { + "group_size": 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": 0.011066652834415436, + "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": 0.007361309137195349, + "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": 0.005892375949770212, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.005631266161799431, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0052945977076888084, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.005054102279245853, + "qparams": { + "group_size": 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": 0.003839333076030016, + "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": 0.003784914268180728, + "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": 0.003379406873136759, + "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": 0.0023885550908744335, + "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": 0.05087171494960785, + "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": 0.05087171494960785, + "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": 0.19332680106163025, + "qparams": { + "group_size": 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": 0.1797168105840683, + "qparams": { + "group_size": 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": 0.17482724785804749, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15702754259109497, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09083499014377594, + "qparams": { + "group_size": 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": 0.08581103384494781, + "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": 0.10354473441839218, + "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": 0.0944814682006836, + "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": 0.09185297787189484, + "qparams": { + "group_size": 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": 0.07988033443689346, + "qparams": { + "group_size": 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": 0.07614948600530624, + "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": 0.052843622863292694, + "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": 0.0451459176838398, + "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": 0.0434429831802845, + "qparams": { + "group_size": 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": 0.04305519908666611, + "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": 0.026449402794241905, + "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": 0.022023968398571014, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021621882915496826, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0197393037378788, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019473515450954437, + "qparams": { + "group_size": 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": 0.013535719364881516, + "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": 0.012917420826852322, + "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": 0.01268838346004486, + "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": 0.007808709051460028, + "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": 0.0944814682006836, + "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": 0.0944814682006836, + "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": 0.133966326713562, + "qparams": { + "group_size": 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": 0.12135528773069382, + "qparams": { + "group_size": 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": 0.11621791124343872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09721004217863083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06276345252990723, + "qparams": { + "group_size": 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": 0.058093488216400146, + "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": 0.0735941231250763, + "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": 0.06698480248451233, + "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": 0.06365254521369934, + "qparams": { + "group_size": 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": 0.0519179105758667, + "qparams": { + "group_size": 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": 0.04692062363028526, + "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": 0.03812123462557793, + "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": 0.03287101909518242, + "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": 0.030928174033761024, + "qparams": { + "group_size": 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": 0.030462995171546936, + "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": 0.019405722618103027, + "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": 0.01717529259622097, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01682095415890217, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.015211491845548153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01491989940404892, + "qparams": { + "group_size": 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": 0.01111825741827488, + "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": 0.012182844802737236, + "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": 0.010544074699282646, + "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": 0.009794878773391247, + "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": 0.09721004217863083, + "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": 0.09721004217863083, + "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": 0.17784526944160461, + "qparams": { + "group_size": 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": 0.1684255301952362, + "qparams": { + "group_size": 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": 0.16543738543987274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15132229030132294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08383212983608246, + "qparams": { + "group_size": 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": 0.08063700050115585, + "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": 0.09260153770446777, + "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": 0.08571983873844147, + "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": 0.08448817580938339, + "qparams": { + "group_size": 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": 0.07592621445655823, + "qparams": { + "group_size": 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": 0.07297347486019135, + "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": 0.047202229499816895, + "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": 0.04109770432114601, + "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": 0.040257927030324936, + "qparams": { + "group_size": 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": 0.04005538672208786, + "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": 0.023591099306941032, + "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": 0.02097744680941105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02076667733490467, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.019406234845519066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01928890496492386, + "qparams": { + "group_size": 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": 0.012544800527393818, + "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": 0.013047964312136173, + "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": 0.012259111739695072, + "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": 0.00919307116419077, + "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": 0.09260153770446777, + "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": 0.09260153770446777, + "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": 0.2223423719406128, + "qparams": { + "group_size": 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": 0.21070358157157898, + "qparams": { + "group_size": 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": 0.2069423943758011, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18937984108924866, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10457035154104233, + "qparams": { + "group_size": 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": 0.10057228058576584, + "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": 0.11539971083402634, + "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": 0.10684482753276825, + "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": 0.10534747689962387, + "qparams": { + "group_size": 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": 0.094641774892807, + "qparams": { + "group_size": 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": 0.09086782485246658, + "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": 0.05854501575231552, + "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": 0.050940852612257004, + "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": 0.049900609999895096, + "qparams": { + "group_size": 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": 0.049647118896245956, + "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": 0.029209380969405174, + "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": 0.025319837033748627, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025054646655917168, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0232800655066967, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02312495745718479, + "qparams": { + "group_size": 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": 0.015055662021040916, + "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": 0.01471659354865551, + "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": 0.014693566597998142, + "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": 0.009074370376765728, + "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": 0.094641774892807, + "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": 0.094641774892807, + "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": 0.19239585101604462, + "qparams": { + "group_size": 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": 0.17185388505458832, + "qparams": { + "group_size": 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": 0.16402795910835266, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14803044497966766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08757711201906204, + "qparams": { + "group_size": 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": 0.08042001724243164, + "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": 0.10342515259981155, + "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": 0.09470394998788834, + "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": 0.09068407118320465, + "qparams": { + "group_size": 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": 0.0766758918762207, + "qparams": { + "group_size": 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": 0.07249182462692261, + "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": 0.05251024663448334, + "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": 0.045533403754234314, + "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": 0.04230061173439026, + "qparams": { + "group_size": 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": 0.04150871932506561, + "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": 0.026467470452189445, + "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": 0.02242555469274521, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02227839082479477, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020332645624876022, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019757360219955444, + "qparams": { + "group_size": 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": 0.01448944490402937, + "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": 0.014820044860243797, + "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": 0.013462618924677372, + "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": 0.010723806917667389, + "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": 0.09470394998788834, + "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": 0.09470394998788834, + "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": 0.06363217532634735, + "qparams": { + "group_size": 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": 0.05892707407474518, + "qparams": { + "group_size": 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": 0.056189607828855515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05053139105439186, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.030007388442754745, + "qparams": { + "group_size": 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": 0.02771054208278656, + "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": 0.03566664457321167, + "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": 0.0329277329146862, + "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": 0.03038564696907997, + "qparams": { + "group_size": 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": 0.02633621357381344, + "qparams": { + "group_size": 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": 0.025194799527525902, + "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": 0.01820097491145134, + "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": 0.015787236392498016, + "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": 0.014456517063081264, + "qparams": { + "group_size": 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": 0.014133526012301445, + "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": 0.009110433980822563, + "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": 0.007532334886491299, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.007297107018530369, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006798441521823406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006592561490833759, + "qparams": { + "group_size": 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": 0.004784095101058483, + "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": 0.004840817302465439, + "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": 0.00433671660721302, + "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": 0.0032734600827097893, + "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": 0.06363217532634735, + "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": 0.06363217532634735, + "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": 0.059195760637521744, + "qparams": { + "group_size": 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": 0.05431491509079933, + "qparams": { + "group_size": 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": 0.051574770361185074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04608842357993126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.027626872062683105, + "qparams": { + "group_size": 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": 0.025263100862503052, + "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": 0.03317869454622269, + "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": 0.030507158488035202, + "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": 0.02803668938577175, + "qparams": { + "group_size": 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": 0.024001561105251312, + "qparams": { + "group_size": 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": 0.02294260449707508, + "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": 0.016815152019262314, + "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": 0.014636521227657795, + "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": 0.013280193321406841, + "qparams": { + "group_size": 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": 0.01294612418860197, + "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": 0.008441926911473274, + "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": 0.006835650652647018, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006590845063328743, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006096404045820236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00588286342099309, + "qparams": { + "group_size": 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": 0.004368219058960676, + "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": 0.004317740444093943, + "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": 0.0039084614254534245, + "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": 0.00277155265212059, + "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": 0.059195760637521744, + "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": 0.059195760637521744, + "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": 0.1848072111606598, + "qparams": { + "group_size": 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": 0.17199720442295074, + "qparams": { + "group_size": 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": 0.1671382337808609, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1508898288011551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08728699386119843, + "qparams": { + "group_size": 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": 0.08230826258659363, + "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": 0.09952636063098907, + "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": 0.0912296324968338, + "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": 0.08828257024288177, + "qparams": { + "group_size": 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": 0.07699698954820633, + "qparams": { + "group_size": 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": 0.0734545961022377, + "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": 0.050846684724092484, + "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": 0.04366182163357735, + "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": 0.041840750724077225, + "qparams": { + "group_size": 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": 0.041381802409887314, + "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": 0.025462601333856583, + "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": 0.02116459794342518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02076183632016182, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.019010039046406746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018713131546974182, + "qparams": { + "group_size": 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": 0.012987049296498299, + "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": 0.012407523579895496, + "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": 0.012181617319583893, + "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": 0.007395725231617689, + "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": 0.09952636063098907, + "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": 0.09952636063098907, + "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": 0.14504635334014893, + "qparams": { + "group_size": 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": 0.12060873210430145, + "qparams": { + "group_size": 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": 0.11162656545639038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09484987705945969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06614255160093307, + "qparams": { + "group_size": 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": 0.057353295385837555, + "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": 0.08162937313318253, + "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": 0.07285197079181671, + "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": 0.06830430775880814, + "qparams": { + "group_size": 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": 0.05067962035536766, + "qparams": { + "group_size": 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": 0.04868386685848236, + "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": 0.04196051508188248, + "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": 0.036188460886478424, + "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": 0.03313780575990677, + "qparams": { + "group_size": 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": 0.03238874301314354, + "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": 0.021794714033603668, + "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": 0.01920366659760475, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01862083561718464, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.016589755192399025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.016127415001392365, + "qparams": { + "group_size": 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": 0.012938368134200573, + "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": 0.014511107467114925, + "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": 0.01196931954473257, + "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": 0.01216744352132082, + "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": 0.09484987705945969, + "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": 0.09484987705945969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "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": 0.16170141100883484, + "qparams": { + "group_size": 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": 0.1532222479581833, + "qparams": { + "group_size": 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": 0.15043121576309204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.13765722513198853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07625453919172287, + "qparams": { + "group_size": 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": 0.07330375164747238, + "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": 0.08431334793567657, + "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": 0.0780705139040947, + "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": 0.07683668285608292, + "qparams": { + "group_size": 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": 0.0690600648522377, + "qparams": { + "group_size": 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": 0.06636674702167511, + "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": 0.04295992851257324, + "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": 0.03742774575948715, + "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": 0.03660286217927933, + "qparams": { + "group_size": 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": 0.036401696503162384, + "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": 0.021474797278642654, + "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": 0.019062353298068047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018866857513785362, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.017644017934799194, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01751592382788658, + "qparams": { + "group_size": 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": 0.011395826935768127, + "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": 0.011842404492199421, + "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": 0.011120679788291454, + "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": 0.008309549652040005, + "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": 0.08431334793567657, + "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": 0.08431334793567657, + "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": 0.21380850672721863, + "qparams": { + "group_size": 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": 0.2026515007019043, + "qparams": { + "group_size": 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": 0.19899991154670715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18219317495822906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1005396619439125, + "qparams": { + "group_size": 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": 0.09666644781827927, + "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": 0.11104641109704971, + "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": 0.10280703753232956, + "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": 0.10131201148033142, + "qparams": { + "group_size": 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": 0.09105246514081955, + "qparams": { + "group_size": 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": 0.08743559569120407, + "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": 0.05631992593407631, + "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": 0.04902185499668121, + "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": 0.047973185777664185, + "qparams": { + "group_size": 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": 0.0477292500436306, + "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": 0.028086503967642784, + "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": 0.024327782914042473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024067677557468414, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02237159200012684, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022213581949472427, + "qparams": { + "group_size": 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": 0.014450819231569767, + "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": 0.014121058396995068, + "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": 0.014087196439504623, + "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": 0.008674263954162598, + "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": 0.1005396619439125, + "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": 0.1005396619439125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.1847352534532547, + "qparams": { + "group_size": 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": 0.16438205540180206, + "qparams": { + "group_size": 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": 0.15654794871807098, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14151312410831451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08378718048334122, + "qparams": { + "group_size": 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": 0.07670244574546814, + "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": 0.1000794991850853, + "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": 0.09124637395143509, + "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": 0.08691442012786865, + "qparams": { + "group_size": 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": 0.07349617779254913, + "qparams": { + "group_size": 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": 0.0698075219988823, + "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": 0.05093444883823395, + "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": 0.04406805709004402, + "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": 0.040712710469961166, + "qparams": { + "group_size": 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": 0.03988366201519966, + "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": 0.025875704362988472, + "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": 0.02201812155544758, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021853674203157425, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020090937614440918, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019515272229909897, + "qparams": { + "group_size": 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": 0.01452578417956829, + "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": 0.015109128318727016, + "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": 0.013463711366057396, + "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": 0.011463078670203686, + "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": 0.1000794991850853, + "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": 0.1000794991850853, + "qparams": { + "group_size": 128, + "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": 0.06924930214881897, + "qparams": { + "group_size": 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": 0.06501559913158417, + "qparams": { + "group_size": 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": 0.06302453577518463, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05723455920815468, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03273377567529678, + "qparams": { + "group_size": 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": 0.030886901542544365, + "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": 0.03812369331717491, + "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": 0.0346611924469471, + "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": 0.03304341807961464, + "qparams": { + "group_size": 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": 0.029230551794171333, + "qparams": { + "group_size": 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": 0.028278935700654984, + "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": 0.019465148448944092, + "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": 0.01662762649357319, + "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": 0.015718093141913414, + "qparams": { + "group_size": 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": 0.01550763938575983, + "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": 0.009745546616613865, + "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": 0.008124032989144325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.007950819097459316, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.007415026426315308, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00727695832028985, + "qparams": { + "group_size": 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": 0.005075038876384497, + "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": 0.005033230409026146, + "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": 0.004690516274422407, + "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": 0.003337458474561572, + "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": 0.06924930214881897, + "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": 0.06924930214881897, + "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": 0.0640471801161766, + "qparams": { + "group_size": 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": 0.05992966890335083, + "qparams": { + "group_size": 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": 0.05769467353820801, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05222161486744881, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.029998894780874252, + "qparams": { + "group_size": 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": 0.028106151148676872, + "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": 0.034927789121866226, + "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": 0.03231247514486313, + "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": 0.030291875824332237, + "qparams": { + "group_size": 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": 0.02668377198278904, + "qparams": { + "group_size": 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": 0.025544075295329094, + "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": 0.017670489847660065, + "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": 0.015428596176207066, + "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": 0.014397739432752132, + "qparams": { + "group_size": 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": 0.01414407603442669, + "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": 0.008836760185658932, + "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": 0.007391204126179218, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.007182599976658821, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006702004000544548, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006533435545861721, + "qparams": { + "group_size": 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": 0.004604910034686327, + "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": 0.004540971014648676, + "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": 0.004263959359377623, + "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": 0.002870909171178937, + "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": 0.0640471801161766, + "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": 0.0640471801161766, + "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": 0.19219212234020233, + "qparams": { + "group_size": 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": 0.18017037212848663, + "qparams": { + "group_size": 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": 0.17581140995025635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15933120250701904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09028946608304977, + "qparams": { + "group_size": 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": 0.08582954853773117, + "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": 0.10187831521034241, + "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": 0.0935860127210617, + "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": 0.09125515073537827, + "qparams": { + "group_size": 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": 0.08047709614038467, + "qparams": { + "group_size": 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": 0.0771080031991005, + "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": 0.05209389328956604, + "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": 0.04480636864900589, + "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": 0.04323870688676834, + "qparams": { + "group_size": 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": 0.04284821078181267, + "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": 0.026043057441711426, + "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": 0.021803118288517, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021449919790029526, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.019749194383621216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01947702281177044, + "qparams": { + "group_size": 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": 0.013279751874506474, + "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": 0.012637311592698097, + "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": 0.012566812336444855, + "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": 0.007455253973603249, + "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": 0.0935860127210617, + "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": 0.0935860127210617, + "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": 0.17586104571819305, + "qparams": { + "group_size": 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": 0.15290004014968872, + "qparams": { + "group_size": 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": 0.14549192786216736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12123329937458038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08192339539527893, + "qparams": { + "group_size": 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": 0.07352468371391296, + "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": 0.09357430785894394, + "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": 0.08589226007461548, + "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": 0.08326775580644608, + "qparams": { + "group_size": 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": 0.06435643136501312, + "qparams": { + "group_size": 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": 0.05828306823968887, + "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": 0.048199113458395004, + "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": 0.04169992730021477, + "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": 0.039911381900310516, + "qparams": { + "group_size": 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": 0.03948913514614105, + "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": 0.024356897920370102, + "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": 0.02155819721519947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021062102168798447, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.018275639042258263, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01799619197845459, + "qparams": { + "group_size": 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": 0.013646794483065605, + "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": 0.01455004420131445, + "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": 0.013106071390211582, + "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": 0.011189599521458149, + "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": 0.09357430785894394, + "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": 0.09357430785894394, + "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": 0.15701982378959656, + "qparams": { + "group_size": 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": 0.1488916426897049, + "qparams": { + "group_size": 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": 0.1462802141904831, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.13395635783672333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07388919591903687, + "qparams": { + "group_size": 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": 0.0710839107632637, + "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": 0.08168688416481018, + "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": 0.07559629529714584, + "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": 0.07442458719015121, + "qparams": { + "group_size": 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": 0.06703877449035645, + "qparams": { + "group_size": 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": 0.06448834389448166, + "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": 0.04156376048922539, + "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": 0.03618870675563812, + "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": 0.03541368991136551, + "qparams": { + "group_size": 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": 0.03522561490535736, + "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": 0.020771626383066177, + "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": 0.018360348418354988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018170811235904694, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.016996517777442932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01688605546951294, + "qparams": { + "group_size": 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": 0.010976199992001057, + "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": 0.011282929219305515, + "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": 0.010713054798543453, + "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": 0.007787321228533983, + "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": 0.08168688416481018, + "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": 0.08168688416481018, + "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": 0.2144310176372528, + "qparams": { + "group_size": 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": 0.20353244245052338, + "qparams": { + "group_size": 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": 0.19996079802513123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18324939906597137, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10067661851644516, + "qparams": { + "group_size": 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": 0.096963070333004, + "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": 0.1112813875079155, + "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": 0.10297822207212448, + "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": 0.10144541412591934, + "qparams": { + "group_size": 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": 0.09143389016389847, + "qparams": { + "group_size": 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": 0.08790703117847443, + "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": 0.056431520730257034, + "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": 0.04908466711640358, + "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": 0.04804820939898491, + "qparams": { + "group_size": 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": 0.04780549556016922, + "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": 0.028147665783762932, + "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": 0.02438529022037983, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024122418835759163, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022469809278845787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022317292168736458, + "qparams": { + "group_size": 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": 0.014509549364447594, + "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": 0.014170781709253788, + "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": 0.014147462323307991, + "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": 0.008733293041586876, + "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": 0.10067661851644516, + "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": 0.10067661851644516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.18781007826328278, + "qparams": { + "group_size": 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": 0.16625292599201202, + "qparams": { + "group_size": 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": 0.15790557861328125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14266212284564972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08507925271987915, + "qparams": { + "group_size": 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": 0.07742633670568466, + "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": 0.10138869285583496, + "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": 0.0930590033531189, + "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": 0.08848907798528671, + "qparams": { + "group_size": 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": 0.07427112758159637, + "qparams": { + "group_size": 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": 0.07055087387561798, + "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": 0.05194247141480446, + "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": 0.04498676583170891, + "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": 0.04134581983089447, + "qparams": { + "group_size": 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": 0.0404910109937191, + "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": 0.026354873552918434, + "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": 0.022416014224290848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02222866751253605, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020396394655108452, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019763462245464325, + "qparams": { + "group_size": 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": 0.01483258418738842, + "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": 0.015450268052518368, + "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": 0.013670460321009159, + "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": 0.011711218394339085, + "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": 0.0930590033531189, + "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": 0.0930590033531189, + "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": 0.06530224531888962, + "qparams": { + "group_size": 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": 0.061213765293359756, + "qparams": { + "group_size": 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": 0.05926017090678215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05326876416802406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.030956203117966652, + "qparams": { + "group_size": 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": 0.029171904549002647, + "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": 0.03643056005239487, + "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": 0.03287642076611519, + "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": 0.03128260374069214, + "qparams": { + "group_size": 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": 0.027319036424160004, + "qparams": { + "group_size": 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": 0.026304103434085846, + "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": 0.01864003762602806, + "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": 0.015833556652069092, + "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": 0.014928760938346386, + "qparams": { + "group_size": 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": 0.014717748388648033, + "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": 0.009361565113067627, + "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": 0.0077432324178516865, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.007574853952974081, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006983062718063593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006845948752015829, + "qparams": { + "group_size": 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": 0.004952270537614822, + "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": 0.004833222832530737, + "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": 0.004649899434298277, + "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": 0.003241799771785736, + "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": 0.06530224531888962, + "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": 0.06530224531888962, + "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": 0.060413893312215805, + "qparams": { + "group_size": 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": 0.05624833703041077, + "qparams": { + "group_size": 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": 0.053905948996543884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.048553820699453354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.02826702781021595, + "qparams": { + "group_size": 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": 0.02629700116813183, + "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": 0.03309180960059166, + "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": 0.030734797939658165, + "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": 0.028581976890563965, + "qparams": { + "group_size": 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": 0.024952998384833336, + "qparams": { + "group_size": 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": 0.02383413538336754, + "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": 0.016766760498285294, + "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": 0.014675626531243324, + "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": 0.013548108749091625, + "qparams": { + "group_size": 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": 0.013281391002237797, + "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": 0.00840702187269926, + "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": 0.006917200051248074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0067085400223731995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006241730414330959, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006052747834473848, + "qparams": { + "group_size": 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": 0.004354104865342379, + "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": 0.004240002483129501, + "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": 0.003988930955529213, + "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": 0.0026132771745324135, + "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": 0.060413893312215805, + "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": 0.060413893312215805, + "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": 0.19191640615463257, + "qparams": { + "group_size": 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": 0.17944513261318207, + "qparams": { + "group_size": 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": 0.17454901337623596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1580078899860382, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09045958518981934, + "qparams": { + "group_size": 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": 0.08575206995010376, + "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": 0.10265247523784637, + "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": 0.09414193779230118, + "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": 0.09141188859939575, + "qparams": { + "group_size": 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": 0.08030728250741959, + "qparams": { + "group_size": 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": 0.07698933780193329, + "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": 0.05247063189744949, + "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": 0.045040473341941833, + "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": 0.04325571656227112, + "qparams": { + "group_size": 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": 0.04285449534654617, + "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": 0.026233509182929993, + "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": 0.02189096435904503, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021497229114174843, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01973499171435833, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019486339762806892, + "qparams": { + "group_size": 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": 0.01343662478029728, + "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": 0.01277762372046709, + "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": 0.01262678112834692, + "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": 0.007597220595926046, + "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": 0.09414193779230118, + "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": 0.09414193779230118, + "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": 0.1754126101732254, + "qparams": { + "group_size": 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": 0.15489035844802856, + "qparams": { + "group_size": 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": 0.14759917557239532, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12063749134540558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0815286636352539, + "qparams": { + "group_size": 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": 0.07393168658018112, + "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": 0.09430252015590668, + "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": 0.08608326315879822, + "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": 0.08292503654956818, + "qparams": { + "group_size": 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": 0.0651492029428482, + "qparams": { + "group_size": 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": 0.05817050859332085, + "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": 0.04844667762517929, + "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": 0.04204407334327698, + "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": 0.03996800258755684, + "qparams": { + "group_size": 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": 0.03947703540325165, + "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": 0.024634599685668945, + "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": 0.022059107199311256, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02157566510140896, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01910487189888954, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018779167905449867, + "qparams": { + "group_size": 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": 0.014079730957746506, + "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": 0.015457410365343094, + "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": 0.013476921245455742, + "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": 0.012331558391451836, + "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": 0.09430252015590668, + "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": 0.09430252015590668, + "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": 0.1736868917942047, + "qparams": { + "group_size": 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": 0.16479675471782684, + "qparams": { + "group_size": 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": 0.1617663949728012, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1483033001422882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08213746547698975, + "qparams": { + "group_size": 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": 0.07904176414012909, + "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": 0.09122809767723083, + "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": 0.08428795635700226, + "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": 0.08279574662446976, + "qparams": { + "group_size": 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": 0.0746823325753212, + "qparams": { + "group_size": 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": 0.07197876274585724, + "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": 0.046870067715644836, + "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": 0.04096505790948868, + "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": 0.04002785682678223, + "qparams": { + "group_size": 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": 0.03980766609311104, + "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": 0.02360626310110092, + "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": 0.021811332553625107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021600397303700447, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020414207130670547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020296426489949226, + "qparams": { + "group_size": 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": 0.013189468532800674, + "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": 0.014826207421720028, + "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": 0.012895660474896431, + "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": 0.01174049824476242, + "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": 0.09122809767723083, + "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": 0.09122809767723083, + "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": 0.2231774628162384, + "qparams": { + "group_size": 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": 0.21178482472896576, + "qparams": { + "group_size": 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": 0.2080504298210144, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1907714456319809, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10495850443840027, + "qparams": { + "group_size": 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": 0.1009974256157875, + "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": 0.11603742092847824, + "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": 0.1074383407831192, + "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": 0.10573398321866989, + "qparams": { + "group_size": 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": 0.095343217253685, + "qparams": { + "group_size": 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": 0.09169292449951172, + "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": 0.059037864208221436, + "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": 0.05128917470574379, + "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": 0.050113461911678314, + "qparams": { + "group_size": 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": 0.04984422028064728, + "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": 0.029480570927262306, + "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": 0.025585191324353218, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02530287578701973, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02361205592751503, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023438073694705963, + "qparams": { + "group_size": 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": 0.015359039418399334, + "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": 0.01512779388576746, + "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": 0.014946999959647655, + "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": 0.009648947045207024, + "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": 0.095343217253685, + "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": 0.095343217253685, + "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": 0.06687191873788834, + "qparams": { + "group_size": 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": 0.06566625833511353, + "qparams": { + "group_size": 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": 0.02449386939406395, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.023733364418148994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.021993432193994522, + "qparams": { + "group_size": 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": 0.006663783453404903, + "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": 0.08473418653011322, + "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": 0.06374572217464447, + "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": 0.022142762318253517, + "qparams": { + "group_size": 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": 0.02161538228392601, + "qparams": { + "group_size": 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": 0.025354888290166855, + "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": 0.025053638964891434, + "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": 0.02109050378203392, + "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": 0.007784164045006037, + "qparams": { + "group_size": 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": 0.004063469357788563, + "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": 0.007870464585721493, + "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": 0.0029439134523272514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0021471299696713686, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0028135417960584164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0019319435814395547, + "qparams": { + "group_size": 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": 0.00262365466915071, + "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": 0.002609583083540201, + "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": 0.0014470306923612952, + "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": 0.001458649872802198, + "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": 0.06687191873788834, + "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": 0.06687191873788834, + "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": 0.08266904205083847, + "qparams": { + "group_size": 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": 0.07740696519613266, + "qparams": { + "group_size": 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": 0.0749121755361557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06783834099769592, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03905746340751648, + "qparams": { + "group_size": 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": 0.03680887073278427, + "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": 0.04508526623249054, + "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": 0.04145297035574913, + "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": 0.03948250040411949, + "qparams": { + "group_size": 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": 0.034757718443870544, + "qparams": { + "group_size": 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": 0.03330203518271446, + "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": 0.02293737605214119, + "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": 0.019849175587296486, + "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": 0.018769696354866028, + "qparams": { + "group_size": 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": 0.018509220331907272, + "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": 0.011472407728433609, + "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": 0.009683056734502316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009469952434301376, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008801317773759365, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008633467368781567, + "qparams": { + "group_size": 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": 0.005999384913593531, + "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": 0.005977231077849865, + "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": 0.005618942901492119, + "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": 0.0039315177127718925, + "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": 0.08266904205083847, + "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": 0.08266904205083847, + "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": 0.0734420046210289, + "qparams": { + "group_size": 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": 0.06842516362667084, + "qparams": { + "group_size": 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": 0.06572634726762772, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05922611430287361, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03427969291806221, + "qparams": { + "group_size": 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": 0.03203120455145836, + "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": 0.04010464996099472, + "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": 0.03721228241920471, + "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": 0.03466765210032463, + "qparams": { + "group_size": 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": 0.03036346659064293, + "qparams": { + "group_size": 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": 0.029036909341812134, + "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": 0.020278414711356163, + "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": 0.017711201682686806, + "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": 0.016442634165287018, + "qparams": { + "group_size": 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": 0.016131669282913208, + "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": 0.010148759000003338, + "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": 0.008396423421800137, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008156055584549904, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.007575327530503273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0073738484643399715, + "qparams": { + "group_size": 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": 0.005237937904894352, + "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": 0.005127892829477787, + "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": 0.004836438223719597, + "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": 0.0031601933296769857, + "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": 0.0734420046210289, + "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": 0.0734420046210289, + "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": 0.2107955813407898, + "qparams": { + "group_size": 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": 0.19779156148433685, + "qparams": { + "group_size": 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": 0.1931201070547104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.17499971389770508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09946900606155396, + "qparams": { + "group_size": 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": 0.09475143998861313, + "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": 0.11193640530109406, + "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": 0.1027812734246254, + "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": 0.1003945842385292, + "qparams": { + "group_size": 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": 0.08851131051778793, + "qparams": { + "group_size": 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": 0.08470264077186584, + "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": 0.057137154042720795, + "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": 0.04908657819032669, + "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": 0.04757729545235634, + "qparams": { + "group_size": 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": 0.04719444364309311, + "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": 0.028578566387295723, + "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": 0.02401733212172985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023638732731342316, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021692000329494476, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02144058234989643, + "qparams": { + "group_size": 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": 0.014564557000994682, + "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": 0.013846280984580517, + "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": 0.013818744570016861, + "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": 0.008118340745568275, + "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": 0.1003945842385292, + "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": 0.09946900606155396, + "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": 0.1676812618970871, + "qparams": { + "group_size": 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": 0.1489666849374771, + "qparams": { + "group_size": 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": 0.14214695990085602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.11760030686855316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07859305292367935, + "qparams": { + "group_size": 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": 0.07177252322435379, + "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": 0.09050844609737396, + "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": 0.08321105688810349, + "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": 0.07976704835891724, + "qparams": { + "group_size": 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": 0.0625682845711708, + "qparams": { + "group_size": 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": 0.05676079913973808, + "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": 0.046637311577796936, + "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": 0.04026325047016144, + "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": 0.038138289004564285, + "qparams": { + "group_size": 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": 0.037630029022693634, + "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": 0.023404357954859734, + "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": 0.02039971947669983, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01994914375245571, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.017359131947159767, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01702042482793331, + "qparams": { + "group_size": 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": 0.012815475463867188, + "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": 0.013579102233052254, + "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": 0.012140639126300812, + "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": 0.010168173350393772, + "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": 0.09050844609737396, + "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": 0.09050844609737396, + "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": 0.17179559171199799, + "qparams": { + "group_size": 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": 0.16268570721149445, + "qparams": { + "group_size": 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": 0.15969669818878174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14605580270290375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08096501231193542, + "qparams": { + "group_size": 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": 0.07778240740299225, + "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": 0.08958977460861206, + "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": 0.08289723843336105, + "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": 0.08160041272640228, + "qparams": { + "group_size": 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": 0.07323217391967773, + "qparams": { + "group_size": 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": 0.07022133469581604, + "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": 0.045485589653253555, + "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": 0.039583686739206314, + "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": 0.03871362283825874, + "qparams": { + "group_size": 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": 0.038508109748363495, + "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": 0.022743800655007362, + "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": 0.01984071545302868, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019625689834356308, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01826312020421028, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01813490502536297, + "qparams": { + "group_size": 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": 0.011902863159775734, + "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": 0.011838053353130817, + "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": 0.01159713976085186, + "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": 0.007725009229034185, + "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": 0.08958977460861206, + "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": 0.08958977460861206, + "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": 0.22883570194244385, + "qparams": { + "group_size": 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": 0.21673734486103058, + "qparams": { + "group_size": 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": 0.21283245086669922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.19473421573638916, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10775887966156006, + "qparams": { + "group_size": 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": 0.10361640900373459, + "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": 0.1189960241317749, + "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": 0.11020869761705399, + "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": 0.10859469324350357, + "qparams": { + "group_size": 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": 0.09753771871328354, + "qparams": { + "group_size": 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": 0.09357636421918869, + "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": 0.06039586290717125, + "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": 0.052558109164237976, + "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": 0.05145672336220741, + "qparams": { + "group_size": 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": 0.051182378083467484, + "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": 0.030123276636004448, + "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": 0.026112155988812447, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025829343125224113, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.023996539413928986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023824434727430344, + "qparams": { + "group_size": 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": 0.015505258925259113, + "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": 0.015185627155005932, + "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": 0.015119774267077446, + "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": 0.00935961864888668, + "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": 0.09753771871328354, + "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": 0.09753771871328354, + "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": 0.21121221780776978, + "qparams": { + "group_size": 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": 0.18787051737308502, + "qparams": { + "group_size": 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": 0.17870648205280304, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16098333895206451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09629806876182556, + "qparams": { + "group_size": 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": 0.08784899860620499, + "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": 0.11491252481937408, + "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": 0.10501362383365631, + "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": 0.09997750073671341, + "qparams": { + "group_size": 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": 0.08386272192001343, + "qparams": { + "group_size": 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": 0.07939571142196655, + "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": 0.058548931032419205, + "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": 0.050520338118076324, + "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": 0.046560138463974, + "qparams": { + "group_size": 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": 0.04557899013161659, + "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": 0.029661666601896286, + "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": 0.024744808673858643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024548830464482307, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02236131951212883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02166442759335041, + "qparams": { + "group_size": 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": 0.016469627618789673, + "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": 0.01648290455341339, + "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": 0.015220741741359234, + "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": 0.011948816478252411, + "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": 0.09997750073671341, + "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": 0.09997750073671341, + "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": 0.06276684999465942, + "qparams": { + "group_size": 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": 0.058464184403419495, + "qparams": { + "group_size": 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": 0.05625017359852791, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05068041756749153, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.02970392070710659, + "qparams": { + "group_size": 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": 0.027723517268896103, + "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": 0.034825220704078674, + "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": 0.03194679692387581, + "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": 0.030061936005949974, + "qparams": { + "group_size": 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": 0.02622825652360916, + "qparams": { + "group_size": 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": 0.025085892528295517, + "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": 0.017765404656529427, + "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": 0.0153916385024786, + "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": 0.014344285242259502, + "qparams": { + "group_size": 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": 0.014099315740168095, + "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": 0.008916596882045269, + "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": 0.0075430311262607574, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.007355190813541412, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006855982355773449, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00670169573277235, + "qparams": { + "group_size": 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": 0.00474441098049283, + "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": 0.00488671101629734, + "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": 0.004389350768178701, + "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": 0.003445509122684598, + "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": 0.06276684999465942, + "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": 0.06276684999465942, + "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": 0.06695246696472168, + "qparams": { + "group_size": 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": 0.061980169266462326, + "qparams": { + "group_size": 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": 0.0591118261218071, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.053055763244628906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.031247958540916443, + "qparams": { + "group_size": 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": 0.028842628002166748, + "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": 0.03709721937775612, + "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": 0.034433621913194656, + "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": 0.03162562474608421, + "qparams": { + "group_size": 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": 0.027467774227261543, + "qparams": { + "group_size": 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": 0.026185432448983192, + "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": 0.01884332485496998, + "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": 0.016432102769613266, + "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": 0.01500733196735382, + "qparams": { + "group_size": 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": 0.014648009091615677, + "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": 0.009418443776667118, + "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": 0.007709364872425795, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.007437008433043957, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.006919617764651775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006690987851470709, + "qparams": { + "group_size": 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": 0.004891769960522652, + "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": 0.00481739966198802, + "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": 0.004421218764036894, + "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": 0.0030223100911825895, + "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": 0.06695246696472168, + "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": 0.06695246696472168, + "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": 0.18569111824035645, + "qparams": { + "group_size": 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": 0.17328238487243652, + "qparams": { + "group_size": 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": 0.16854318976402283, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1520717740058899, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08745691925287247, + "qparams": { + "group_size": 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": 0.08265865594148636, + "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": 0.10008155554533005, + "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": 0.0912722498178482, + "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": 0.08836516737937927, + "qparams": { + "group_size": 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": 0.07741019874811172, + "qparams": { + "group_size": 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": 0.0741593986749649, + "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": 0.05099130794405937, + "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": 0.0436827689409256, + "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": 0.04187130928039551, + "qparams": { + "group_size": 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": 0.04143600910902023, + "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": 0.025485944002866745, + "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": 0.02116730995476246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020773258060216904, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01904914714396, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01877644658088684, + "qparams": { + "group_size": 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": 0.01300766784697771, + "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": 0.01233834307640791, + "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": 0.01216043159365654, + "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": 0.007284805178642273, + "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": 0.10008155554533005, + "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": 0.10008155554533005, + "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": 0.16790251433849335, + "qparams": { + "group_size": 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": 0.14981476962566376, + "qparams": { + "group_size": 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": 0.13902486860752106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12076453119516373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07813026010990143, + "qparams": { + "group_size": 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": 0.0695946142077446, + "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": 0.10357356816530228, + "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": 0.08998193591833115, + "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": 0.08010682463645935, + "qparams": { + "group_size": 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": 0.06571415066719055, + "qparams": { + "group_size": 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": 0.06273153424263, + "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": 0.05246410518884659, + "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": 0.04401130974292755, + "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": 0.03870353102684021, + "qparams": { + "group_size": 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": 0.037339214235544205, + "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": 0.026834197342395782, + "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": 0.02182639203965664, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020950792357325554, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.019645949825644493, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018810920417308807, + "qparams": { + "group_size": 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": 0.015394818969070911, + "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": 0.01625104621052742, + "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": 0.01349497027695179, + "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": 0.012970777228474617, + "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": 0.08998193591833115, + "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": 0.08998193591833115, + "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": 0.16167670488357544, + "qparams": { + "group_size": 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": 0.15314197540283203, + "qparams": { + "group_size": 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": 0.15029291808605194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1374172568321228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07627785205841064, + "qparams": { + "group_size": 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": 0.07329763472080231, + "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": 0.0843687355518341, + "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": 0.07811476290225983, + "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": 0.07687941193580627, + "qparams": { + "group_size": 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": 0.06899375468492508, + "qparams": { + "group_size": 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": 0.06618465483188629, + "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": 0.04294431954622269, + "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": 0.037383802235126495, + "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": 0.03656291589140892, + "qparams": { + "group_size": 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": 0.03636579588055611, + "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": 0.021443290635943413, + "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": 0.018895164132118225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018691204488277435, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.017428265884518623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01730693131685257, + "qparams": { + "group_size": 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": 0.011277123354375362, + "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": 0.011516297236084938, + "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": 0.010999973863363266, + "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": 0.007818536832928658, + "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": 0.0843687355518341, + "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": 0.0843687355518341, + "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": 0.22747616469860077, + "qparams": { + "group_size": 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": 0.21557828783988953, + "qparams": { + "group_size": 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": 0.21171273291110992, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1936284750699997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10713944584131241, + "qparams": { + "group_size": 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": 0.10303860902786255, + "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": 0.11853176355361938, + "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": 0.10960954427719116, + "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": 0.10799113661050797, + "qparams": { + "group_size": 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": 0.0970073789358139, + "qparams": { + "group_size": 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": 0.09303154051303864, + "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": 0.06009085103869438, + "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": 0.052275314927101135, + "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": 0.051160138100385666, + "qparams": { + "group_size": 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": 0.05089738592505455, + "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": 0.029991818591952324, + "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": 0.02596890740096569, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025691917166113853, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.023858698084950447, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023690126836299896, + "qparams": { + "group_size": 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": 0.015474128536880016, + "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": 0.015100670978426933, + "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": 0.015088086947798729, + "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": 0.00930631160736084, + "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": 0.0970073789358139, + "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": 0.0970073789358139, + "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": 0.2069949358701706, + "qparams": { + "group_size": 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": 0.1838647574186325, + "qparams": { + "group_size": 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": 0.17520582675933838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1574256718158722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09442301094532013, + "qparams": { + "group_size": 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": 0.08629771322011948, + "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": 0.11160974949598312, + "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": 0.10219074040651321, + "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": 0.09799021482467651, + "qparams": { + "group_size": 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": 0.08189615607261658, + "qparams": { + "group_size": 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": 0.07733440399169922, + "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": 0.05678524076938629, + "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": 0.04914960637688637, + "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": 0.045650798827409744, + "qparams": { + "group_size": 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": 0.04480098560452461, + "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": 0.028582725673913956, + "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": 0.024262042716145515, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024123016744852066, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021862316876649857, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02124343439936638, + "qparams": { + "group_size": 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": 0.0156419575214386, + "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": 0.016118774190545082, + "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": 0.014534842222929, + "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": 0.011755766347050667, + "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": 0.09799021482467651, + "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": 0.09799021482467651, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.08665397763252258, + "qparams": { + "group_size": 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": 0.08058727532625198, + "qparams": { + "group_size": 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": 0.07766162604093552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06994768977165222, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.040810465812683105, + "qparams": { + "group_size": 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": 0.0381714291870594, + "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": 0.047311607748270035, + "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": 0.04355889931321144, + "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": 0.041260477155447006, + "qparams": { + "group_size": 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": 0.035997163504362106, + "qparams": { + "group_size": 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": 0.034346070140600204, + "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": 0.024126356467604637, + "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": 0.020923767238855362, + "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": 0.01965327374637127, + "qparams": { + "group_size": 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": 0.019346807152032852, + "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": 0.012077965773642063, + "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": 0.010252992622554302, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010006807744503021, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.009287090040743351, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009091444313526154, + "qparams": { + "group_size": 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": 0.006366427056491375, + "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": 0.006516278721392155, + "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": 0.00592468585819006, + "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": 0.00447640847414732, + "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": 0.08665397763252258, + "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": 0.08665397763252258, + "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": 0.07917021214962006, + "qparams": { + "group_size": 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": 0.07306735962629318, + "qparams": { + "group_size": 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": 0.07023624330759048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06283146888017654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.036761827766895294, + "qparams": { + "group_size": 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": 0.034151140600442886, + "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": 0.0431152768433094, + "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": 0.03960902616381645, + "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": 0.037250760942697525, + "qparams": { + "group_size": 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": 0.03221866488456726, + "qparams": { + "group_size": 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": 0.03067130781710148, + "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": 0.021790023893117905, + "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": 0.018941478803753853, + "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": 0.017678052186965942, + "qparams": { + "group_size": 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": 0.017352625727653503, + "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": 0.01092124730348587, + "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": 0.009115751832723618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008871159516274929, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008191286586225033, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.007991334423422813, + "qparams": { + "group_size": 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": 0.005715231876820326, + "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": 0.00569146778434515, + "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": 0.005267086438834667, + "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": 0.003733595134690404, + "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": 0.07917021214962006, + "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": 0.07917021214962006, + "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": 0.19757108390331268, + "qparams": { + "group_size": 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": 0.18496735394001007, + "qparams": { + "group_size": 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": 0.1805475801229477, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16323411464691162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09323163330554962, + "qparams": { + "group_size": 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": 0.08849205821752548, + "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": 0.10543529689311981, + "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": 0.09658734500408173, + "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": 0.094153493642807, + "qparams": { + "group_size": 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": 0.08294734358787537, + "qparams": { + "group_size": 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": 0.07916256785392761, + "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": 0.05383843556046486, + "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": 0.04618819057941437, + "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": 0.04463263973593712, + "qparams": { + "group_size": 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": 0.0442369244992733, + "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": 0.02688654512166977, + "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": 0.02262670174241066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022266335785388947, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020448897033929825, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020194994285702705, + "qparams": { + "group_size": 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": 0.013793897815048695, + "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": 0.013203497976064682, + "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": 0.01309640146791935, + "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": 0.007966837845742702, + "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": 0.09658734500408173, + "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": 0.09658734500408173, + "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": 0.1902107298374176, + "qparams": { + "group_size": 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": 0.17329081892967224, + "qparams": { + "group_size": 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": 0.16675356030464172, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14496766030788422, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08919721096754074, + "qparams": { + "group_size": 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": 0.0824376717209816, + "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": 0.1035228744149208, + "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": 0.09425022453069687, + "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": 0.09046769887208939, + "qparams": { + "group_size": 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": 0.07571642100811005, + "qparams": { + "group_size": 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": 0.07004882395267487, + "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": 0.05317646637558937, + "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": 0.045536432415246964, + "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": 0.04322068393230438, + "qparams": { + "group_size": 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": 0.04268180951476097, + "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": 0.026856224983930588, + "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": 0.022986050695180893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022485898807644844, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02037590742111206, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02001815475523472, + "qparams": { + "group_size": 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": 0.014861071482300758, + "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": 0.015124514698982239, + "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": 0.014128811657428741, + "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": 0.011156896129250526, + "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": 0.09425022453069687, + "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": 0.09425022453069687, + "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": 0.1495862603187561, + "qparams": { + "group_size": 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": 0.14162053167819977, + "qparams": { + "group_size": 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": 0.13887272775173187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12679637968540192, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0706920251250267, + "qparams": { + "group_size": 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": 0.06785007566213608, + "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": 0.07843375205993652, + "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": 0.07258766144514084, + "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": 0.07127183675765991, + "qparams": { + "group_size": 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": 0.06386103481054306, + "qparams": { + "group_size": 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": 0.06122541427612305, + "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": 0.03994067758321762, + "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": 0.034746650606393814, + "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": 0.0338958278298378, + "qparams": { + "group_size": 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": 0.03370172902941704, + "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": 0.01996190845966339, + "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": 0.017532803118228912, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017332127317786217, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01615416817367077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01602894999086857, + "qparams": { + "group_size": 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": 0.010528281331062317, + "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": 0.01071527786552906, + "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": 0.010250061750411987, + "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": 0.007289642468094826, + "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": 0.07843375205993652, + "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": 0.07843375205993652, + "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": 0.22961868345737457, + "qparams": { + "group_size": 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": 0.2174111306667328, + "qparams": { + "group_size": 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": 0.21343742311000824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1950129270553589, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10836626589298248, + "qparams": { + "group_size": 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": 0.10407274961471558, + "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": 0.119963139295578, + "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": 0.11096496134996414, + "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": 0.10921965539455414, + "qparams": { + "group_size": 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": 0.09792236238718033, + "qparams": { + "group_size": 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": 0.09391985088586807, + "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": 0.06091638654470444, + "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": 0.052971016615629196, + "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": 0.051769230514764786, + "qparams": { + "group_size": 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": 0.05149069055914879, + "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": 0.030388709157705307, + "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": 0.026312589645385742, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02601720206439495, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02415316551923752, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023973707109689713, + "qparams": { + "group_size": 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": 0.015665916725993156, + "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": 0.01536975521594286, + "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": 0.015246391296386719, + "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": 0.00955415889620781, + "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": 0.09792236238718033, + "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": 0.09792236238718033, + "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": 0.21390388906002045, + "qparams": { + "group_size": 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": 0.18920844793319702, + "qparams": { + "group_size": 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": 0.17956161499023438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16110286116600037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09730523079633713, + "qparams": { + "group_size": 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": 0.08841663599014282, + "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": 0.11606252193450928, + "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": 0.10627289116382599, + "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": 0.10120326280593872, + "qparams": { + "group_size": 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": 0.08414394408464432, + "qparams": { + "group_size": 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": 0.0795554369688034, + "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": 0.05923011153936386, + "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": 0.051158465445041656, + "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": 0.047083210200071335, + "qparams": { + "group_size": 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": 0.046100419014692307, + "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": 0.029819298535585403, + "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": 0.02512894570827484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024944385513663292, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022616831585764885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021895967423915863, + "qparams": { + "group_size": 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": 0.016380133107304573, + "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": 0.016874175518751144, + "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": 0.015116707421839237, + "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": 0.012363213114440441, + "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": 0.09730523079633713, + "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": 0.09730523079633713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.075381338596344, + "qparams": { + "group_size": 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": 0.0707254558801651, + "qparams": { + "group_size": 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": 0.06854380667209625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06208033114671707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03563571348786354, + "qparams": { + "group_size": 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": 0.0336531400680542, + "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": 0.040992219001054764, + "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": 0.03772583231329918, + "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": 0.03598697483539581, + "qparams": { + "group_size": 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": 0.03177417814731598, + "qparams": { + "group_size": 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": 0.030411750078201294, + "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": 0.020855162292718887, + "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": 0.018073730170726776, + "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": 0.017125818878412247, + "qparams": { + "group_size": 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": 0.016893070191144943, + "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": 0.010442202910780907, + "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": 0.008856154978275299, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008675757795572281, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008070986717939377, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.007925890386104584, + "qparams": { + "group_size": 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": 0.005485539324581623, + "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": 0.0054986304603517056, + "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": 0.005149567499756813, + "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": 0.003669964149594307, + "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": 0.075381338596344, + "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": 0.075381338596344, + "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": 0.06303980201482773, + "qparams": { + "group_size": 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": 0.058983199298381805, + "qparams": { + "group_size": 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": 0.05647600442171097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05105385184288025, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.02943398989737034, + "qparams": { + "group_size": 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": 0.02746904455125332, + "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": 0.03451812267303467, + "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": 0.032247669994831085, + "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": 0.029786178842186928, + "qparams": { + "group_size": 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": 0.02622714266180992, + "qparams": { + "group_size": 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": 0.025052402168512344, + "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": 0.017423050478100777, + "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": 0.01533021591603756, + "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": 0.014118148013949394, + "qparams": { + "group_size": 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": 0.01381501741707325, + "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": 0.00870693288743496, + "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": 0.007206709124147892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006989327259361744, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0065360539592802525, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.006345245987176895, + "qparams": { + "group_size": 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": 0.004509935155510902, + "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": 0.0044160643592476845, + "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": 0.004144322592765093, + "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": 0.0027165657375007868, + "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": 0.06303980201482773, + "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": 0.06303980201482773, + "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": 0.2147434800863266, + "qparams": { + "group_size": 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": 0.20148661732673645, + "qparams": { + "group_size": 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": 0.19700202345848083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.17792868614196777, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10100708901882172, + "qparams": { + "group_size": 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": 0.0964118093252182, + "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": 0.11371703445911407, + "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": 0.10414965450763702, + "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": 0.10198774933815002, + "qparams": { + "group_size": 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": 0.0899863913655281, + "qparams": { + "group_size": 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": 0.0860394760966301, + "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": 0.05806173384189606, + "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": 0.049772538244724274, + "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": 0.04832038655877113, + "qparams": { + "group_size": 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": 0.04795176163315773, + "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": 0.028996266424655914, + "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": 0.024406878277659416, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024060839787125587, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02209652028977871, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02187015675008297, + "qparams": { + "group_size": 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": 0.014824619516730309, + "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": 0.014105224050581455, + "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": 0.014074415899813175, + "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": 0.008348739705979824, + "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": 0.0899863913655281, + "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": 0.0899863913655281, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "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": 0.1842668056488037, + "qparams": { + "group_size": 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": 0.1668962687253952, + "qparams": { + "group_size": 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": 0.15839490294456482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1374722123146057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08482007682323456, + "qparams": { + "group_size": 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": 0.0776895210146904, + "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": 0.10336523503065109, + "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": 0.09426452219486237, + "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": 0.08754362165927887, + "qparams": { + "group_size": 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": 0.07325133681297302, + "qparams": { + "group_size": 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": 0.06842678785324097, + "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": 0.05337778106331825, + "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": 0.045864518731832504, + "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": 0.041494011878967285, + "qparams": { + "group_size": 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": 0.04036073014140129, + "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": 0.027066338807344437, + "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": 0.02264494076371193, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02197178825736046, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020497838035225868, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019813159480690956, + "qparams": { + "group_size": 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": 0.014849638566374779, + "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": 0.015919450670480728, + "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": 0.013244839385151863, + "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": 0.012173091992735863, + "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": 0.09426452219486237, + "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": 0.09426452219486237, + "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": 0.13959264755249023, + "qparams": { + "group_size": 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": 0.13215933740139008, + "qparams": { + "group_size": 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": 0.12956230342388153, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.11834874749183655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06597382575273514, + "qparams": { + "group_size": 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": 0.06331004202365875, + "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": 0.0732097178697586, + "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": 0.06775537133216858, + "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": 0.06650378555059433, + "qparams": { + "group_size": 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": 0.05961114540696144, + "qparams": { + "group_size": 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": 0.05718255415558815, + "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": 0.03730858117341995, + "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": 0.03251507878303528, + "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": 0.03170317783951759, + "qparams": { + "group_size": 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": 0.03151721507310867, + "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": 0.018662791699171066, + "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": 0.016532480716705322, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01634969934821129, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.015272634103894234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.015157359652221203, + "qparams": { + "group_size": 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": 0.009923982433974743, + "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": 0.010310005396604538, + "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": 0.009660215117037296, + "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": 0.007252123206853867, + "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": 0.0732097178697586, + "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": 0.0732097178697586, + "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": 0.21326906979084015, + "qparams": { + "group_size": 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": 0.20207099616527557, + "qparams": { + "group_size": 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": 0.1982627660036087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18130739033222198, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10066026449203491, + "qparams": { + "group_size": 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": 0.09663118422031403, + "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": 0.11152602732181549, + "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": 0.10314479470252991, + "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": 0.10144156962633133, + "qparams": { + "group_size": 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": 0.09098052233457565, + "qparams": { + "group_size": 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": 0.08730718493461609, + "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": 0.05664372071623802, + "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": 0.04922991618514061, + "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": 0.04809218645095825, + "qparams": { + "group_size": 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": 0.04782276973128319, + "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": 0.02826741337776184, + "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": 0.02448064647614956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02419973351061344, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02248035930097103, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022310808300971985, + "qparams": { + "group_size": 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": 0.014599023386836052, + "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": 0.014352119527757168, + "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": 0.014196661300957203, + "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": 0.008996288292109966, + "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": 0.10066026449203491, + "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": 0.10066026449203491, + "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": 0.20744115114212036, + "qparams": { + "group_size": 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": 0.1832682490348816, + "qparams": { + "group_size": 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": 0.17409390211105347, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15549910068511963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09471409767866135, + "qparams": { + "group_size": 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": 0.08592601865530014, + "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": 0.1125248372554779, + "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": 0.10291484743356705, + "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": 0.09831835329532623, + "qparams": { + "group_size": 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": 0.08145899325609207, + "qparams": { + "group_size": 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": 0.07680297642946243, + "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": 0.057710085064172745, + "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": 0.04970979318022728, + "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": 0.045994460582733154, + "qparams": { + "group_size": 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": 0.04509911686182022, + "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": 0.029155591502785683, + "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": 0.024787817150354385, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024625275284051895, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022280337288975716, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021625909954309464, + "qparams": { + "group_size": 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": 0.016250604763627052, + "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": 0.016895519569516182, + "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": 0.015104871243238449, + "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": 0.01272725872695446, + "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": 0.09831835329532623, + "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": 0.09831835329532623, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.09428081661462784, + "qparams": { + "group_size": 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": 0.08812755346298218, + "qparams": { + "group_size": 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": 0.08543763309717178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07714319229125977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.044310614466667175, + "qparams": { + "group_size": 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": 0.04179393872618675, + "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": 0.050705742090940475, + "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": 0.04673769325017929, + "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": 0.04475056380033493, + "qparams": { + "group_size": 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": 0.039329662919044495, + "qparams": { + "group_size": 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": 0.037560902535915375, + "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": 0.025810481980443, + "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": 0.022400304675102234, + "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": 0.02128531038761139, + "qparams": { + "group_size": 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": 0.021020440384745598, + "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": 0.012924780137836933, + "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": 0.011033954098820686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010812792927026749, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.010034168139100075, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009866880252957344, + "qparams": { + "group_size": 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": 0.006789931561797857, + "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": 0.006878342479467392, + "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": 0.006396276876330376, + "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": 0.004651154391467571, + "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": 0.09428081661462784, + "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": 0.09428081661462784, + "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": 0.08086983859539032, + "qparams": { + "group_size": 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": 0.0754125565290451, + "qparams": { + "group_size": 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": 0.07269727438688278, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06528473645448685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03763752803206444, + "qparams": { + "group_size": 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": 0.03527332469820976, + "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": 0.04345342889428139, + "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": 0.04014822095632553, + "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": 0.03803163021802902, + "qparams": { + "group_size": 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": 0.03321990370750427, + "qparams": { + "group_size": 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": 0.03169688582420349, + "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": 0.0220147967338562, + "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": 0.019167518243193626, + "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": 0.018019158393144608, + "qparams": { + "group_size": 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": 0.017756517976522446, + "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": 0.011003296822309494, + "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": 0.009247949346899986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009031252935528755, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008342968299984932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008173280395567417, + "qparams": { + "group_size": 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": 0.005704366602003574, + "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": 0.00565132312476635, + "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": 0.005315563641488552, + "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": 0.003625683020800352, + "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": 0.08086983859539032, + "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": 0.08086983859539032, + "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": 0.19198311865329742, + "qparams": { + "group_size": 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": 0.18025442957878113, + "qparams": { + "group_size": 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": 0.1760113537311554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15937045216560364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0902705043554306, + "qparams": { + "group_size": 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": 0.08589151501655579, + "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": 0.1017298772931099, + "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": 0.09329230338335037, + "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": 0.09109035134315491, + "qparams": { + "group_size": 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": 0.08045057952404022, + "qparams": { + "group_size": 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": 0.076930932700634, + "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": 0.05171317607164383, + "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": 0.04450878128409386, + "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": 0.04313018172979355, + "qparams": { + "group_size": 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": 0.042820267379283905, + "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": 0.025814559310674667, + "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": 0.021811239421367645, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02149568498134613, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01974223554134369, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019545644521713257, + "qparams": { + "group_size": 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": 0.013188645243644714, + "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": 0.012607075273990631, + "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": 0.01260188315063715, + "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": 0.007484250236302614, + "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": 0.09329230338335037, + "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": 0.09329230338335037, + "qparams": { + "group_size": 32, + "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": 0.1891206055879593, + "qparams": { + "group_size": 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": 0.1693446934223175, + "qparams": { + "group_size": 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": 0.1621810644865036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.13973382115364075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08807027339935303, + "qparams": { + "group_size": 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": 0.08082073926925659, + "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": 0.10224805772304535, + "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": 0.09346716850996017, + "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": 0.08984871953725815, + "qparams": { + "group_size": 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": 0.07326386123895645, + "qparams": { + "group_size": 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": 0.06792985647916794, + "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": 0.05264708399772644, + "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": 0.045153986662626266, + "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": 0.04268959164619446, + "qparams": { + "group_size": 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": 0.04210754856467247, + "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": 0.026479976251721382, + "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": 0.02270887978374958, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022185076028108597, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.019846711307764053, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0194765105843544, + "qparams": { + "group_size": 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": 0.014320923015475273, + "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": 0.014980753883719444, + "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": 0.013547860085964203, + "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": 0.011044167913496494, + "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": 0.09346716850996017, + "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": 0.09346716850996017, + "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": 0.144900843501091, + "qparams": { + "group_size": 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": 0.13716089725494385, + "qparams": { + "group_size": 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": 0.13455750048160553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12288840115070343, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06846165657043457, + "qparams": { + "group_size": 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": 0.06571017950773239, + "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": 0.07594018429517746, + "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": 0.07026843726634979, + "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": 0.06898348778486252, + "qparams": { + "group_size": 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": 0.06185333430767059, + "qparams": { + "group_size": 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": 0.059344302862882614, + "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": 0.038691021502017975, + "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": 0.033694397658109665, + "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": 0.03287120535969734, + "qparams": { + "group_size": 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": 0.03267392888665199, + "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": 0.019340401515364647, + "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": 0.017093762755393982, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01690448448061943, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.015776941552758217, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.015652675181627274, + "qparams": { + "group_size": 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": 0.010240819305181503, + "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": 0.01058407686650753, + "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": 0.009963308461010456, + "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": 0.007369760423898697, + "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": 0.07594018429517746, + "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": 0.07594018429517746, + "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": 0.21355611085891724, + "qparams": { + "group_size": 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": 0.20222008228302002, + "qparams": { + "group_size": 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": 0.19850844144821167, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18144789338111877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10076083987951279, + "qparams": { + "group_size": 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": 0.0967569574713707, + "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": 0.11165963858366013, + "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": 0.10325980186462402, + "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": 0.10156383365392685, + "qparams": { + "group_size": 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": 0.09108293056488037, + "qparams": { + "group_size": 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": 0.08732233196496964, + "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": 0.05668678879737854, + "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": 0.04927149787545204, + "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": 0.04815002903342247, + "qparams": { + "group_size": 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": 0.04787237569689751, + "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": 0.028281809762120247, + "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": 0.024493269622325897, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02421603910624981, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022485138848423958, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022315416485071182, + "qparams": { + "group_size": 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": 0.014578157104551792, + "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": 0.014337336644530296, + "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": 0.014176690950989723, + "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": 0.008939152583479881, + "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": 0.09108293056488037, + "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": 0.09108293056488037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "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": 0.21308863162994385, + "qparams": { + "group_size": 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": 0.18819715082645416, + "qparams": { + "group_size": 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": 0.1786557137966156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15920524299144745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09747754782438278, + "qparams": { + "group_size": 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": 0.08839623630046844, + "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": 0.11545637995004654, + "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": 0.1059737280011177, + "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": 0.10116951167583466, + "qparams": { + "group_size": 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": 0.08346572518348694, + "qparams": { + "group_size": 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": 0.07846791297197342, + "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": 0.05910862982273102, + "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": 0.051159102469682693, + "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": 0.04734005406498909, + "qparams": { + "group_size": 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": 0.04639142379164696, + "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": 0.029776278883218765, + "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": 0.025459222495555878, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025296781212091446, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022818120196461678, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02213602513074875, + "qparams": { + "group_size": 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": 0.016377631574869156, + "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": 0.01732276752591133, + "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": 0.015168257057666779, + "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": 0.013006001710891724, + "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": 0.09747754782438278, + "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": 0.09747754782438278, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.09833546727895737, + "qparams": { + "group_size": 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": 0.09198594093322754, + "qparams": { + "group_size": 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": 0.08913304656744003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08045848459005356, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0463445670902729, + "qparams": { + "group_size": 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": 0.043656542897224426, + "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": 0.053316231817007065, + "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": 0.04901242256164551, + "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": 0.04679499939084053, + "qparams": { + "group_size": 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": 0.04113125056028366, + "qparams": { + "group_size": 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": 0.039336174726486206, + "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": 0.02719067595899105, + "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": 0.023540357127785683, + "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": 0.022285431623458862, + "qparams": { + "group_size": 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": 0.021985001862049103, + "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": 0.013599376194179058, + "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": 0.011534573510289192, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011296097189188004, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.010489803738892078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010300066322088242, + "qparams": { + "group_size": 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": 0.007114013656973839, + "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": 0.007193717174232006, + "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": 0.006653827615082264, + "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": 0.004835034720599651, + "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": 0.09833546727895737, + "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": 0.09833546727895737, + "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": 0.0844482034444809, + "qparams": { + "group_size": 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": 0.07855480909347534, + "qparams": { + "group_size": 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": 0.07556550204753876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06801024079322815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03935929015278816, + "qparams": { + "group_size": 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": 0.0367916077375412, + "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": 0.04599481076002121, + "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": 0.0423131100833416, + "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": 0.03981553763151169, + "qparams": { + "group_size": 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": 0.03473202884197235, + "qparams": { + "group_size": 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": 0.03317060321569443, + "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": 0.02324945293366909, + "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": 0.020213846117258072, + "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": 0.018878964707255363, + "qparams": { + "group_size": 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": 0.01854400709271431, + "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": 0.011640843003988266, + "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": 0.009692846797406673, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009441756643354893, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008749452419579029, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008533565327525139, + "qparams": { + "group_size": 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": 0.006046482361853123, + "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": 0.00597523944452405, + "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": 0.005578037351369858, + "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": 0.0038128942251205444, + "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": 0.0844482034444809, + "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": 0.0844482034444809, + "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": 0.19693994522094727, + "qparams": { + "group_size": 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": 0.18518990278244019, + "qparams": { + "group_size": 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": 0.18063956499099731, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16336339712142944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09276802837848663, + "qparams": { + "group_size": 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": 0.08822476118803024, + "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": 0.10456129163503647, + "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": 0.09609514474868774, + "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": 0.09360258281230927, + "qparams": { + "group_size": 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": 0.0825905054807663, + "qparams": { + "group_size": 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": 0.07898172736167908, + "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": 0.053252965211868286, + "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": 0.04586854949593544, + "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": 0.044316310435533524, + "qparams": { + "group_size": 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": 0.04393596202135086, + "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": 0.026532387360930443, + "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": 0.02235657349228859, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022007398307323456, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020232677459716797, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019985785707831383, + "qparams": { + "group_size": 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": 0.013579132035374641, + "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": 0.012915484607219696, + "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": 0.01289868913590908, + "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": 0.0075670816004276276, + "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": 0.09609514474868774, + "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": 0.09609514474868774, + "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": 0.1905553936958313, + "qparams": { + "group_size": 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": 0.17459556460380554, + "qparams": { + "group_size": 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": 0.1685703843832016, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1473315954208374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08973664790391922, + "qparams": { + "group_size": 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": 0.08370890468358994, + "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": 0.10404671728610992, + "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": 0.09398351609706879, + "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": 0.0910593643784523, + "qparams": { + "group_size": 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": 0.07654707878828049, + "qparams": { + "group_size": 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": 0.0722445547580719, + "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": 0.0537470206618309, + "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": 0.045309070497751236, + "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": 0.043349236249923706, + "qparams": { + "group_size": 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": 0.04288657754659653, + "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": 0.02687930129468441, + "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": 0.022696036845445633, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022264225408434868, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020069284364581108, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019775398075580597, + "qparams": { + "group_size": 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": 0.014379305765032768, + "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": 0.014424871653318405, + "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": 0.013735754415392876, + "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": 0.010213221423327923, + "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": 0.09398351609706879, + "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": 0.09398351609706879, + "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": 0.15110760927200317, + "qparams": { + "group_size": 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": 0.14302273094654083, + "qparams": { + "group_size": 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": 0.1402103453874588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1282252073287964, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07143787294626236, + "qparams": { + "group_size": 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": 0.06853955239057541, + "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": 0.07921707630157471, + "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": 0.07333895564079285, + "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": 0.07200999557971954, + "qparams": { + "group_size": 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": 0.06453415751457214, + "qparams": { + "group_size": 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": 0.06192222610116005, + "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": 0.040379106998443604, + "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": 0.03517880290746689, + "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": 0.03431335836648941, + "qparams": { + "group_size": 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": 0.03411354497075081, + "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": 0.020206041634082794, + "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": 0.017858270555734634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017656218260526657, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01648145355284214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.016357429325580597, + "qparams": { + "group_size": 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": 0.010719461366534233, + "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": 0.011076707392930984, + "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": 0.01043827086687088, + "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": 0.007730002049356699, + "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": 0.07921707630157471, + "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": 0.07921707630157471, + "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": 0.2176220417022705, + "qparams": { + "group_size": 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": 0.20594662427902222, + "qparams": { + "group_size": 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": 0.20210668444633484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1846916526556015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10262001305818558, + "qparams": { + "group_size": 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": 0.09856078028678894, + "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": 0.11360548436641693, + "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": 0.1051349937915802, + "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": 0.1034679263830185, + "qparams": { + "group_size": 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": 0.09271428734064102, + "qparams": { + "group_size": 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": 0.0888996422290802, + "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": 0.05774281173944473, + "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": 0.05018442124128342, + "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": 0.049046840518713, + "qparams": { + "group_size": 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": 0.048780594021081924, + "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": 0.028811924159526825, + "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": 0.02496621012687683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024689393118023872, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022916335612535477, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022742627188563347, + "qparams": { + "group_size": 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": 0.014886045828461647, + "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": 0.014639501459896564, + "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": 0.014482485130429268, + "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": 0.009168840013444424, + "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": 0.09271428734064102, + "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": 0.09271428734064102, + "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": 0.22032059729099274, + "qparams": { + "group_size": 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": 0.194514662027359, + "qparams": { + "group_size": 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": 0.18483898043632507, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16384883224964142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10101848840713501, + "qparams": { + "group_size": 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": 0.09174495935440063, + "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": 0.11895987391471863, + "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": 0.10946399718523026, + "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": 0.10476159304380417, + "qparams": { + "group_size": 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": 0.08619280904531479, + "qparams": { + "group_size": 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": 0.08078467100858688, + "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": 0.061059076339006424, + "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": 0.0528457835316658, + "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": 0.04899773374199867, + "qparams": { + "group_size": 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": 0.04805997759103775, + "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": 0.03070640005171299, + "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": 0.026291174814105034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02613731101155281, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.023474065586924553, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022796908393502235, + "qparams": { + "group_size": 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": 0.016856860369443893, + "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": 0.01776054874062538, + "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": 0.015652954578399658, + "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": 0.013277627527713776, + "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": 0.08619280904531479, + "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": 0.08619280904531479, + "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": 0.10477587580680847, + "qparams": { + "group_size": 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": 0.09821201860904694, + "qparams": { + "group_size": 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": 0.09551329910755157, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08638375997543335, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04941976070404053, + "qparams": { + "group_size": 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": 0.04682707041501999, + "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": 0.056111257523298264, + "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": 0.051746778190135956, + "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": 0.04986964166164398, + "qparams": { + "group_size": 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": 0.043989960104227066, + "qparams": { + "group_size": 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": 0.04198439046740532, + "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": 0.028576917946338654, + "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": 0.024806812405586243, + "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": 0.02375427447259426, + "qparams": { + "group_size": 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": 0.023501528427004814, + "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": 0.014286563731729984, + "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": 0.012318872846662998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012103009037673473, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011230334639549255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011068650521337986, + "qparams": { + "group_size": 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": 0.007497093174606562, + "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": 0.007660347502678633, + "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": 0.007125133648514748, + "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": 0.005221320781856775, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09821201860904694, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09821201860904694, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "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": 0.07889635115861893, + "qparams": { + "group_size": 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": 0.07394738495349884, + "qparams": { + "group_size": 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": 0.071345254778862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06436461210250854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.036918602883815765, + "qparams": { + "group_size": 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": 0.03465817868709564, + "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": 0.04271111264824867, + "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": 0.0395476333796978, + "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": 0.037275075912475586, + "qparams": { + "group_size": 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": 0.032792333513498306, + "qparams": { + "group_size": 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": 0.0313587561249733, + "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": 0.021610025316476822, + "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": 0.01889452151954174, + "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": 0.017688479274511337, + "qparams": { + "group_size": 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": 0.01738400012254715, + "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": 0.010814945213496685, + "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": 0.00903137307614088, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00879922416061163, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008190767839550972, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008006490767002106, + "qparams": { + "group_size": 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": 0.005607505328953266, + "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": 0.00549325393512845, + "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": 0.00522862933576107, + "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": 0.003415366169065237, + "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": 0.07889635115861893, + "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": 0.07889635115861893, + "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": 0.22186121344566345, + "qparams": { + "group_size": 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": 0.20841968059539795, + "qparams": { + "group_size": 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": 0.20402206480503082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18448618054389954, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10441377758979797, + "qparams": { + "group_size": 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": 0.09964663535356522, + "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": 0.11704734712839127, + "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": 0.10746407508850098, + "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": 0.10531269013881683, + "qparams": { + "group_size": 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": 0.09307066351175308, + "qparams": { + "group_size": 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": 0.08897213637828827, + "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": 0.05963284522294998, + "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": 0.0513782873749733, + "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": 0.04988913610577583, + "qparams": { + "group_size": 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": 0.04953913018107414, + "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": 0.029822388663887978, + "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": 0.025187762454152107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02483118698000908, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022798649966716766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022572163492441177, + "qparams": { + "group_size": 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": 0.015208613127470016, + "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": 0.014455387368798256, + "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": 0.014529107138514519, + "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": 0.00849059782922268, + "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": 0.09307066351175308, + "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": 0.09307066351175308, + "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": 0.19524063169956207, + "qparams": { + "group_size": 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": 0.18003849685192108, + "qparams": { + "group_size": 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": 0.17356713116168976, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15500497817993164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09128493070602417, + "qparams": { + "group_size": 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": 0.0849342942237854, + "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": 0.105738565325737, + "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": 0.09713254868984222, + "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": 0.0924021378159523, + "qparams": { + "group_size": 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": 0.07939358055591583, + "qparams": { + "group_size": 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": 0.07516349852085114, + "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": 0.05418655276298523, + "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": 0.04687803238630295, + "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": 0.04418731853365898, + "qparams": { + "group_size": 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": 0.043529920279979706, + "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": 0.027260206639766693, + "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": 0.02351858839392662, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023004107177257538, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021213334053754807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020805522799491882, + "qparams": { + "group_size": 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": 0.01482318714261055, + "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": 0.015551706776022911, + "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": 0.013989854604005814, + "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": 0.011465595103800297, + "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": 0.09713254868984222, + "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": 0.09713254868984222, + "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": 0.1517738550901413, + "qparams": { + "group_size": 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": 0.14356157183647156, + "qparams": { + "group_size": 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": 0.14067798852920532, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12832459807395935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07177199423313141, + "qparams": { + "group_size": 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": 0.06881404668092728, + "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": 0.07963760197162628, + "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": 0.07373139262199402, + "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": 0.07236064970493317, + "qparams": { + "group_size": 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": 0.06469037383794785, + "qparams": { + "group_size": 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": 0.06197255849838257, + "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": 0.04060628265142441, + "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": 0.03536967188119888, + "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": 0.03448890149593353, + "qparams": { + "group_size": 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": 0.03427799046039581, + "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": 0.02029096521437168, + "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": 0.01793721504509449, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017736047506332397, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.016531435772776604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.016405394300818443, + "qparams": { + "group_size": 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": 0.01074032112956047, + "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": 0.011131775565445423, + "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": 0.010444273240864277, + "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": 0.007751336321234703, + "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": 0.07963760197162628, + "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": 0.07963760197162628, + "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": 0.21809813380241394, + "qparams": { + "group_size": 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": 0.2062433958053589, + "qparams": { + "group_size": 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": 0.20226216316223145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1845794916152954, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10294041037559509, + "qparams": { + "group_size": 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": 0.09872482717037201, + "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": 0.11409078538417816, + "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": 0.10551359504461288, + "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": 0.10374531149864197, + "qparams": { + "group_size": 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": 0.09277242422103882, + "qparams": { + "group_size": 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": 0.08883334696292877, + "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": 0.05793101713061333, + "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": 0.050366442650556564, + "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": 0.049199722707271576, + "qparams": { + "group_size": 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": 0.04892006143927574, + "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": 0.02893482893705368, + "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": 0.025061696767807007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02476591430604458, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02296532317996025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022790320217609406, + "qparams": { + "group_size": 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": 0.014988317154347897, + "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": 0.014719785191118717, + "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": 0.014579241164028645, + "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": 0.00925934687256813, + "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": 0.09277242422103882, + "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": 0.09277242422103882, + "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": 0.22211812436580658, + "qparams": { + "group_size": 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": 0.19608092308044434, + "qparams": { + "group_size": 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": 0.1861141175031662, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16541935503482819, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10169035196304321, + "qparams": { + "group_size": 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": 0.09226900339126587, + "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": 0.12019886076450348, + "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": 0.11047877371311188, + "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": 0.10567692667245865, + "qparams": { + "group_size": 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": 0.08688123524188995, + "qparams": { + "group_size": 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": 0.0815257653594017, + "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": 0.06151135638356209, + "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": 0.05318428575992584, + "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": 0.04920820891857147, + "qparams": { + "group_size": 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": 0.04824388399720192, + "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": 0.030859442427754402, + "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": 0.026185115799307823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02602054551243782, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02333107776939869, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022615013644099236, + "qparams": { + "group_size": 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": 0.016691340133547783, + "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": 0.017442582175135612, + "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": 0.015437825582921505, + "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": 0.012725746259093285, + "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": 0.08688123524188995, + "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": 0.08688123524188995, + "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": 0.09993904829025269, + "qparams": { + "group_size": 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": 0.09384772926568985, + "qparams": { + "group_size": 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": 0.09121047705411911, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08261115849018097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04720991104841232, + "qparams": { + "group_size": 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": 0.04473026469349861, + "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": 0.05372919887304306, + "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": 0.04944181069731712, + "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": 0.04765229672193527, + "qparams": { + "group_size": 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": 0.042068976908922195, + "qparams": { + "group_size": 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": 0.04028049483895302, + "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": 0.02738386020064354, + "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": 0.023691941052675247, + "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": 0.022658472880721092, + "qparams": { + "group_size": 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": 0.022413326427340508, + "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": 0.013684900477528572, + "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": 0.011665083467960358, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011461859568953514, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.010626422241330147, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01046743057668209, + "qparams": { + "group_size": 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": 0.007123761810362339, + "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": 0.007135803811252117, + "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": 0.006736076436936855, + "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": 0.004700296092778444, + "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": 0.09993904829025269, + "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": 0.09993904829025269, + "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": 0.0812377780675888, + "qparams": { + "group_size": 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": 0.07611995190382004, + "qparams": { + "group_size": 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": 0.07362474501132965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.066557377576828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03808431699872017, + "qparams": { + "group_size": 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": 0.035874053835868835, + "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": 0.04386438429355621, + "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": 0.04061845317482948, + "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": 0.03846045583486557, + "qparams": { + "group_size": 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": 0.033958833664655685, + "qparams": { + "group_size": 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": 0.03243827447295189, + "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": 0.02225257083773613, + "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": 0.01944310963153839, + "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": 0.01826157234609127, + "qparams": { + "group_size": 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": 0.017972785979509354, + "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": 0.011126330122351646, + "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": 0.009322171099483967, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00910265278071165, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008466450497508049, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008287256583571434, + "qparams": { + "group_size": 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": 0.00578311225399375, + "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": 0.005652140360325575, + "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": 0.005411728285253048, + "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": 0.0035250794608145952, + "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": 0.0812377780675888, + "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": 0.0812377780675888, + "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": 0.21731366217136383, + "qparams": { + "group_size": 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": 0.20388059318065643, + "qparams": { + "group_size": 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": 0.1992667019367218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18037383258342743, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10227850079536438, + "qparams": { + "group_size": 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": 0.09741146117448807, + "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": 0.11509773135185242, + "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": 0.10545965284109116, + "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": 0.10323020070791245, + "qparams": { + "group_size": 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": 0.09098260849714279, + "qparams": { + "group_size": 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": 0.08698946982622147, + "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": 0.05874493345618248, + "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": 0.050344400107860565, + "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": 0.04883960261940956, + "qparams": { + "group_size": 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": 0.048500824719667435, + "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": 0.029294468462467194, + "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": 0.024715984240174294, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02436256967484951, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022361664101481438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022114301100373268, + "qparams": { + "group_size": 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": 0.014979900792241096, + "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": 0.014279330149292946, + "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": 0.01427040807902813, + "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": 0.008458860218524933, + "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": 0.09098260849714279, + "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": 0.09098260849714279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "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": 0.20078282058238983, + "qparams": { + "group_size": 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": 0.181603342294693, + "qparams": { + "group_size": 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": 0.17449471354484558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1560690999031067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09337318688631058, + "qparams": { + "group_size": 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": 0.08587025105953217, + "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": 0.10825467109680176, + "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": 0.09892939776182175, + "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": 0.09511178731918335, + "qparams": { + "group_size": 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": 0.0798853412270546, + "qparams": { + "group_size": 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": 0.0758529081940651, + "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": 0.05544132739305496, + "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": 0.04765474796295166, + "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": 0.045073252171278, + "qparams": { + "group_size": 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": 0.04447818174958229, + "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": 0.02786659449338913, + "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": 0.02374238520860672, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023187750950455666, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021102771162986755, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020713157951831818, + "qparams": { + "group_size": 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": 0.015004576183855534, + "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": 0.015343056991696358, + "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": 0.014188205823302269, + "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": 0.010984793305397034, + "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": 0.09892939776182175, + "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": 0.09892939776182175, + "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": 0.15931786596775055, + "qparams": { + "group_size": 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": 0.15058544278144836, + "qparams": { + "group_size": 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": 0.14745351672172546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.13452431559562683, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07533092051744461, + "qparams": { + "group_size": 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": 0.07214973866939545, + "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": 0.08383159339427948, + "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": 0.07753988355398178, + "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": 0.07595232129096985, + "qparams": { + "group_size": 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": 0.06790710240602493, + "qparams": { + "group_size": 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": 0.06507125496864319, + "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": 0.042775947600603104, + "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": 0.037232063710689545, + "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": 0.036229878664016724, + "qparams": { + "group_size": 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": 0.03600098565220833, + "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": 0.021396387368440628, + "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": 0.018941203132271767, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018712393939495087, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01746930368244648, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017322080209851265, + "qparams": { + "group_size": 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": 0.011396755464375019, + "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": 0.011889331042766571, + "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": 0.011070186272263527, + "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": 0.008417217060923576, + "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": 0.08383159339427948, + "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": 0.08383159339427948, + "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": 0.21346890926361084, + "qparams": { + "group_size": 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": 0.20190396904945374, + "qparams": { + "group_size": 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": 0.19786584377288818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1804925799369812, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10071361809968948, + "qparams": { + "group_size": 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": 0.0966077521443367, + "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": 0.11194897443056107, + "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": 0.10339730232954025, + "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": 0.1015484631061554, + "qparams": { + "group_size": 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": 0.0907912403345108, + "qparams": { + "group_size": 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": 0.08695319294929504, + "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": 0.05683515593409538, + "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": 0.04937398433685303, + "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": 0.04816119000315666, + "qparams": { + "group_size": 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": 0.04786624386906624, + "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": 0.028392480686306953, + "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": 0.024570513516664505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024274887517094612, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0225306935608387, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022346436977386475, + "qparams": { + "group_size": 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": 0.014735245145857334, + "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": 0.014520546421408653, + "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": 0.014300545677542686, + "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": 0.009214297868311405, + "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": 0.0907912403345108, + "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": 0.0907912403345108, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "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": 0.23056726157665253, + "qparams": { + "group_size": 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": 0.2034999281167984, + "qparams": { + "group_size": 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": 0.19314110279083252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1704827845096588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10606829077005386, + "qparams": { + "group_size": 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": 0.09622157365083694, + "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": 0.12548722326755524, + "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": 0.11508040875196457, + "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": 0.10993070155382156, + "qparams": { + "group_size": 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": 0.09015139937400818, + "qparams": { + "group_size": 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": 0.08420387655496597, + "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": 0.06434176862239838, + "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": 0.055456630885601044, + "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": 0.05135449022054672, + "qparams": { + "group_size": 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": 0.050343647599220276, + "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": 0.03234028071165085, + "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": 0.027317171916365623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.027133874595165253, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02426149696111679, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023511603474617004, + "qparams": { + "group_size": 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": 0.01753738708794117, + "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": 0.01818813942372799, + "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": 0.01622990146279335, + "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": 0.01327589526772499, + "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": 0.09015139937400818, + "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": 0.09015139937400818, + "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": 0.1019616574048996, + "qparams": { + "group_size": 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": 0.09562373161315918, + "qparams": { + "group_size": 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": 0.09271672368049622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08395030349493027, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04816918820142746, + "qparams": { + "group_size": 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": 0.045520707964897156, + "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": 0.05525597929954529, + "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": 0.050830770283937454, + "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": 0.04865147918462753, + "qparams": { + "group_size": 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": 0.042966704815626144, + "qparams": { + "group_size": 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": 0.04112065210938454, + "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": 0.028227636590600014, + "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": 0.02448166161775589, + "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": 0.02321671135723591, + "qparams": { + "group_size": 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": 0.0229085274040699, + "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": 0.014142490923404694, + "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": 0.012103864923119545, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011860968545079231, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011058208532631397, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010869920253753662, + "qparams": { + "group_size": 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": 0.007434974890202284, + "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": 0.007654071785509586, + "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": 0.006975496653467417, + "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": 0.0052991569973528385, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09562373161315918, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09562373161315918, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "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": 0.07907064259052277, + "qparams": { + "group_size": 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": 0.07388097047805786, + "qparams": { + "group_size": 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": 0.07086560130119324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06380784511566162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03691355139017105, + "qparams": { + "group_size": 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": 0.03444663807749748, + "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": 0.04323629289865494, + "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": 0.03999756649136543, + "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": 0.037291545420885086, + "qparams": { + "group_size": 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": 0.032743725925683975, + "qparams": { + "group_size": 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": 0.03128869831562042, + "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": 0.02189631015062332, + "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": 0.019155437126755714, + "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": 0.017723387107253075, + "qparams": { + "group_size": 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": 0.01740650087594986, + "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": 0.010915393941104412, + "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": 0.009139791131019592, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00889111589640379, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008298429660499096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008094380609691143, + "qparams": { + "group_size": 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": 0.005704984068870544, + "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": 0.0057245115749537945, + "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": 0.005267266184091568, + "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": 0.003748697228729725, + "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": 0.07907064259052277, + "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": 0.07907064259052277, + "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": 0.19905714690685272, + "qparams": { + "group_size": 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": 0.18677428364753723, + "qparams": { + "group_size": 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": 0.1821570247411728, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16464363038539886, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09358149766921997, + "qparams": { + "group_size": 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": 0.08883815258741379, + "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": 0.10643023997545242, + "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": 0.0974026769399643, + "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": 0.09442444145679474, + "qparams": { + "group_size": 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": 0.08347104489803314, + "qparams": { + "group_size": 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": 0.07991907000541687, + "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": 0.054316446185112, + "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": 0.046549465507268906, + "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": 0.04470446705818176, + "qparams": { + "group_size": 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": 0.044226858764886856, + "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": 0.027149168774485588, + "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": 0.022628778591752052, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022234106436371803, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020494556054472923, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0202060304582119, + "qparams": { + "group_size": 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": 0.013868567533791065, + "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": 0.013215761631727219, + "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": 0.013051403686404228, + "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": 0.007829085923731327, + "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": 0.0974026769399643, + "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": 0.0974026769399643, + "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": 0.19083863496780396, + "qparams": { + "group_size": 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": 0.1758832186460495, + "qparams": { + "group_size": 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": 0.1664188653230667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14667488634586334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08960355818271637, + "qparams": { + "group_size": 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": 0.08197075873613358, + "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": 0.11046110093593597, + "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": 0.1003316342830658, + "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": 0.09071194380521774, + "qparams": { + "group_size": 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": 0.07776617258787155, + "qparams": { + "group_size": 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": 0.07392829656600952, + "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": 0.056910496205091476, + "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": 0.04880671203136444, + "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": 0.043864693492650986, + "qparams": { + "group_size": 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": 0.042629461735486984, + "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": 0.028820673003792763, + "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": 0.024008313193917274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023203589022159576, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02181822806596756, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021077042445540428, + "qparams": { + "group_size": 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": 0.01598157361149788, + "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": 0.016998033970594406, + "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": 0.014433733187615871, + "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": 0.013084972277283669, + "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": 0.1003316342830658, + "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": 0.1003316342830658, + "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": 0.15604957938194275, + "qparams": { + "group_size": 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": 0.14750970900058746, + "qparams": { + "group_size": 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": 0.14443965256214142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.131752148270607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07385715842247009, + "qparams": { + "group_size": 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": 0.07074234634637833, + "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": 0.0822659358382225, + "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": 0.07607810944318771, + "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": 0.07444814592599869, + "qparams": { + "group_size": 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": 0.066591776907444, + "qparams": { + "group_size": 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": 0.06382233649492264, + "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": 0.0419786274433136, + "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": 0.036551278084516525, + "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": 0.03555026650428772, + "qparams": { + "group_size": 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": 0.035314008593559265, + "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": 0.02101599983870983, + "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": 0.01859566569328308, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018374070525169373, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.0171648096293211, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017011288553476334, + "qparams": { + "group_size": 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": 0.011217287741601467, + "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": 0.011692322790622711, + "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": 0.010883569717407227, + "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": 0.00831298902630806, + "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": 0.0822659358382225, + "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": 0.0822659358382225, + "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": 0.19995200634002686, + "qparams": { + "group_size": 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": 0.18904614448547363, + "qparams": { + "group_size": 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": 0.18529640138149261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16908089816570282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09437782317399979, + "qparams": { + "group_size": 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": 0.0904536098241806, + "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": 0.10493574291467667, + "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": 0.09696867316961288, + "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": 0.09515465795993805, + "qparams": { + "group_size": 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": 0.0850844457745552, + "qparams": { + "group_size": 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": 0.08150272071361542, + "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": 0.05330547317862511, + "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": 0.046303872019052505, + "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": 0.04512237757444382, + "qparams": { + "group_size": 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": 0.04483786225318909, + "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": 0.026597239077091217, + "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": 0.023018071427941322, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02273857407271862, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021113237366080284, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020933859050273895, + "qparams": { + "group_size": 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": 0.01376052014529705, + "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": 0.013609464280307293, + "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": 0.01333999540656805, + "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": 0.008636032231152058, + "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": 0.09696867316961288, + "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": 0.09696867316961288, + "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": 0.21282830834388733, + "qparams": { + "group_size": 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": 0.1870473325252533, + "qparams": { + "group_size": 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": 0.1772051453590393, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1563023030757904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09774762392044067, + "qparams": { + "group_size": 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": 0.08845996856689453, + "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": 0.11623837053775787, + "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": 0.10638559609651566, + "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": 0.1015242338180542, + "qparams": { + "group_size": 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": 0.08290354907512665, + "qparams": { + "group_size": 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": 0.07744193077087402, + "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": 0.05960383266210556, + "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": 0.05133972316980362, + "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": 0.04743663966655731, + "qparams": { + "group_size": 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": 0.04647029563784599, + "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": 0.03001539595425129, + "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": 0.02542657032608986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025253646075725555, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02261102758347988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021907510235905647, + "qparams": { + "group_size": 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": 0.016431931406259537, + "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": 0.017182622104883194, + "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": 0.015194534324109554, + "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": 0.012778905220329762, + "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": 0.09774762392044067, + "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": 0.09774762392044067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.11027169227600098, + "qparams": { + "group_size": 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": 0.10386344790458679, + "qparams": { + "group_size": 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": 0.10121691226959229, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09198140352964401, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05203356593847275, + "qparams": { + "group_size": 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": 0.04952266067266464, + "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": 0.05882485955953598, + "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": 0.05425531789660454, + "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": 0.05248184874653816, + "qparams": { + "group_size": 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": 0.04663422703742981, + "qparams": { + "group_size": 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": 0.04465092346072197, + "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": 0.029921598732471466, + "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": 0.025963224470615387, + "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": 0.024961896240711212, + "qparams": { + "group_size": 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": 0.024723082780838013, + "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": 0.014943144284188747, + "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": 0.01284368522465229, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012632455676794052, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011743462644517422, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011590543203055859, + "qparams": { + "group_size": 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": 0.00777446711435914, + "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": 0.007822835817933083, + "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": 0.0074117244221270084, + "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": 0.005154157057404518, + "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": 0.09198140352964401, + "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": 0.09198140352964401, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "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": 0.07922159135341644, + "qparams": { + "group_size": 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": 0.07446582615375519, + "qparams": { + "group_size": 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": 0.07198548316955566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06512545049190521, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03703520447015762, + "qparams": { + "group_size": 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": 0.034952376037836075, + "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": 0.0425800122320652, + "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": 0.0395355224609375, + "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": 0.03741820901632309, + "qparams": { + "group_size": 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": 0.03314660117030144, + "qparams": { + "group_size": 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": 0.0316554494202137, + "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": 0.02155390754342079, + "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": 0.018874017521739006, + "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": 0.017742207273840904, + "qparams": { + "group_size": 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": 0.0174818467348814, + "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": 0.010802081786096096, + "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": 0.009080723859369755, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008866866119205952, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008273420855402946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008100594393908978, + "qparams": { + "group_size": 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": 0.005626029800623655, + "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": 0.005522027146071196, + "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": 0.005270445719361305, + "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": 0.003464453388005495, + "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": 0.07922159135341644, + "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": 0.07922159135341644, + "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": 0.22218404710292816, + "qparams": { + "group_size": 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": 0.2089986354112625, + "qparams": { + "group_size": 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": 0.20427612960338593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18521980941295624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10433827340602875, + "qparams": { + "group_size": 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": 0.09944011270999908, + "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": 0.11687999218702316, + "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": 0.10781943798065186, + "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": 0.10528731346130371, + "qparams": { + "group_size": 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": 0.09343605488538742, + "qparams": { + "group_size": 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": 0.08921739459037781, + "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": 0.059558723121881485, + "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": 0.051515303552150726, + "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": 0.049921877682209015, + "qparams": { + "group_size": 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": 0.04949602112174034, + "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": 0.029825793579220772, + "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": 0.02519277296960354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02479492872953415, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022897006943821907, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022632649168372154, + "qparams": { + "group_size": 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": 0.015244630165398121, + "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": 0.014609619043767452, + "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": 0.014564204961061478, + "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": 0.008676936849951744, + "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": 0.09343605488538742, + "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": 0.09343605488538742, + "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": 0.20069077610969543, + "qparams": { + "group_size": 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": 0.18375803530216217, + "qparams": { + "group_size": 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": 0.17864125967025757, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15642932057380676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0942496731877327, + "qparams": { + "group_size": 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": 0.08879335969686508, + "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": 0.10731649398803711, + "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": 0.09712737053632736, + "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": 0.09532156586647034, + "qparams": { + "group_size": 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": 0.08058265596628189, + "qparams": { + "group_size": 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": 0.076666921377182, + "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": 0.055585313588380814, + "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": 0.04706951975822449, + "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": 0.04573674872517586, + "qparams": { + "group_size": 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": 0.0454174242913723, + "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": 0.02799222059547901, + "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": 0.02432537078857422, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02397269941866398, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02167423442006111, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02147812210023403, + "qparams": { + "group_size": 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": 0.015222706831991673, + "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": 0.015831351280212402, + "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": 0.014810471795499325, + "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": 0.011824860237538815, + "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": 0.09712737053632736, + "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": 0.09712737053632736, + "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": 0.1630307137966156, + "qparams": { + "group_size": 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": 0.15415342152118683, + "qparams": { + "group_size": 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": 0.15109848976135254, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.13790807127952576, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07706303149461746, + "qparams": { + "group_size": 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": 0.07389333099126816, + "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": 0.08554292470216751, + "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": 0.07919315248727798, + "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": 0.0776977390050888, + "qparams": { + "group_size": 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": 0.06953664124011993, + "qparams": { + "group_size": 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": 0.06667882204055786, + "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": 0.043568044900894165, + "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": 0.037982046604156494, + "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": 0.03701404482126236, + "qparams": { + "group_size": 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": 0.0367860272526741, + "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": 0.021782249212265015, + "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": 0.01924177072942257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01901857927441597, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01773446798324585, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017591219395399094, + "qparams": { + "group_size": 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": 0.011504009366035461, + "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": 0.011914417147636414, + "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": 0.011185811832547188, + "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": 0.008269252255558968, + "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": 0.08554292470216751, + "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": 0.08554292470216751, + "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": 0.20037108659744263, + "qparams": { + "group_size": 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": 0.1895851194858551, + "qparams": { + "group_size": 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": 0.18594108521938324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1698181927204132, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09455308318138123, + "qparams": { + "group_size": 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": 0.0907192975282669, + "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": 0.10486059635877609, + "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": 0.09695818275213242, + "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": 0.09531879425048828, + "qparams": { + "group_size": 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": 0.08533202111721039, + "qparams": { + "group_size": 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": 0.08176280558109283, + "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": 0.05330134183168411, + "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": 0.04631368815898895, + "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": 0.04521624743938446, + "qparams": { + "group_size": 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": 0.0449465811252594, + "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": 0.026615053415298462, + "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": 0.023075895383954048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022808516398072243, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02117958292365074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021011531352996826, + "qparams": { + "group_size": 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": 0.013807003386318684, + "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": 0.013636492192745209, + "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": 0.013418454676866531, + "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": 0.008673310279846191, + "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": 0.09695818275213242, + "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": 0.09695818275213242, + "qparams": { + "group_size": 32, + "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": 0.2189919650554657, + "qparams": { + "group_size": 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": 0.19140858948230743, + "qparams": { + "group_size": 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": 0.18081587553024292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15883006155490875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10001828521490097, + "qparams": { + "group_size": 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": 0.09008976817131042, + "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": 0.11943656206130981, + "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": 0.10936568677425385, + "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": 0.10411593317985535, + "qparams": { + "group_size": 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": 0.08435147255659103, + "qparams": { + "group_size": 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": 0.0785827562212944, + "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": 0.06106432527303696, + "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": 0.05268734693527222, + "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": 0.04846205189824104, + "qparams": { + "group_size": 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": 0.04741811007261276, + "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": 0.030807701870799065, + "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": 0.025879554450511932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025697335600852966, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022885607555508614, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022118285298347473, + "qparams": { + "group_size": 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": 0.016927925869822502, + "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": 0.01739821769297123, + "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": 0.015602671541273594, + "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": 0.0127756018191576, + "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": 0.10001828521490097, + "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": 0.10001828521490097, + "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": 0.11612964421510696, + "qparams": { + "group_size": 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": 0.10949279367923737, + "qparams": { + "group_size": 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": 0.10676540434360504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09705603122711182, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.054848626255989075, + "qparams": { + "group_size": 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": 0.052220553159713745, + "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": 0.06193756312131882, + "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": 0.057109374552965164, + "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": 0.05530955642461777, + "qparams": { + "group_size": 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": 0.04919460415840149, + "qparams": { + "group_size": 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": 0.0470757856965065, + "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": 0.03147774934768677, + "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": 0.027306346222758293, + "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": 0.02630431577563286, + "qparams": { + "group_size": 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": 0.02606327086687088, + "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": 0.01573566533625126, + "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": 0.013544938527047634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.013332679867744446, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.012395869940519333, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012242169119417667, + "qparams": { + "group_size": 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": 0.008199609816074371, + "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": 0.008254586718976498, + "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": 0.007828286848962307, + "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": 0.005457818508148193, + "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": 0.09705603122711182, + "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": 0.09705603122711182, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "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": 0.08434664458036423, + "qparams": { + "group_size": 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": 0.07922658324241638, + "qparams": { + "group_size": 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": 0.07656145095825195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06926465779542923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03950829058885574, + "qparams": { + "group_size": 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": 0.03718272224068642, + "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": 0.04543584585189819, + "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": 0.0423208586871624, + "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": 0.03988184407353401, + "qparams": { + "group_size": 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": 0.03534339740872383, + "qparams": { + "group_size": 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": 0.033826012164354324, + "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": 0.023020293563604355, + "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": 0.020241539925336838, + "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": 0.018945880234241486, + "qparams": { + "group_size": 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": 0.018626514822244644, + "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": 0.011533280834555626, + "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": 0.00967491790652275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00943689327687025, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.008831098675727844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008612955920398235, + "qparams": { + "group_size": 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": 0.005986908450722694, + "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": 0.005899399984627962, + "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": 0.00558314798399806, + "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": 0.003692385507747531, + "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": 0.08434664458036423, + "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": 0.08434664458036423, + "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": 0.23978304862976074, + "qparams": { + "group_size": 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": 0.22554393112659454, + "qparams": { + "group_size": 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": 0.22004446387290955, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.19982822239398956, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.11320151388645172, + "qparams": { + "group_size": 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": 0.10758964717388153, + "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": 0.1277812421321869, + "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": 0.11723688244819641, + "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": 0.1142178326845169, + "qparams": { + "group_size": 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": 0.10108901560306549, + "qparams": { + "group_size": 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": 0.09694457799196243, + "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": 0.06528890877962112, + "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": 0.05615438148379326, + "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": 0.05410117655992508, + "qparams": { + "group_size": 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": 0.05360768735408783, + "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": 0.032615385949611664, + "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": 0.027278007939457893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02681616321206093, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02475259266793728, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024435540661215782, + "qparams": { + "group_size": 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": 0.01665244810283184, + "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": 0.015795312821865082, + "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": 0.015785152092576027, + "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": 0.00922475941479206, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option_max": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09694457799196243, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + "best_option": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09694457799196243, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "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": 0.19432717561721802, + "qparams": { + "group_size": 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": 0.17510150372982025, + "qparams": { + "group_size": 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": 0.16779263317584991, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1470871865749359, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08990851789712906, + "qparams": { + "group_size": 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": 0.08245036751031876, + "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": 0.10486184805631638, + "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": 0.09587609767913818, + "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": 0.09170608967542648, + "qparams": { + "group_size": 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": 0.07640980184078217, + "qparams": { + "group_size": 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": 0.07055754214525223, + "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": 0.05361603945493698, + "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": 0.04614312946796417, + "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": 0.04347781836986542, + "qparams": { + "group_size": 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": 0.04281550645828247, + "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": 0.026983430609107018, + "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": 0.023004859685897827, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022433551028370857, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02038651332259178, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019976753741502762, + "qparams": { + "group_size": 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": 0.014624121598899364, + "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": 0.015050698071718216, + "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": 0.013775626197457314, + "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": 0.010924587957561016, + "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": 0.09587609767913818, + "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": 0.09587609767913818, + "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": 0.1664293259382248, + "qparams": { + "group_size": 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": 0.15754885971546173, + "qparams": { + "group_size": 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": 0.15446172654628754, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14119473099708557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07874593883752823, + "qparams": { + "group_size": 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": 0.07558402419090271, + "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": 0.08729955554008484, + "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": 0.08080915361642838, + "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": 0.07938142865896225, + "qparams": { + "group_size": 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": 0.07112617045640945, + "qparams": { + "group_size": 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": 0.06822869181632996, + "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": 0.044525668025016785, + "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": 0.03875160589814186, + "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": 0.037817005068063736, + "qparams": { + "group_size": 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": 0.03760359063744545, + "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": 0.02224624902009964, + "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": 0.0196355152875185, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019416101276874542, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01811259053647518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01796838827431202, + "qparams": { + "group_size": 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": 0.011761834844946861, + "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": 0.012121965177357197, + "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": 0.01145325880497694, + "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": 0.008388254791498184, + "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": 0.08729955554008484, + "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": 0.08729955554008484, + "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": 0.2128899246454239, + "qparams": { + "group_size": 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": 0.2015012502670288, + "qparams": { + "group_size": 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": 0.1977207213640213, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18073402345180511, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10050547122955322, + "qparams": { + "group_size": 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": 0.09650831669569016, + "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": 0.11115317791700363, + "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": 0.10296712070703506, + "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": 0.10130485147237778, + "qparams": { + "group_size": 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": 0.09078151732683182, + "qparams": { + "group_size": 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": 0.08701907843351364, + "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": 0.056490641087293625, + "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": 0.04914660379290581, + "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": 0.04803474619984627, + "qparams": { + "group_size": 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": 0.047778911888599396, + "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": 0.028185246512293816, + "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": 0.024441052228212357, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02416842058300972, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022436654195189476, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022263456135988235, + "qparams": { + "group_size": 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": 0.01455482468008995, + "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": 0.0143263665959239, + "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": 0.014169364236295223, + "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": 0.008967066183686256, + "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": 0.10050547122955322, + "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": 0.10050547122955322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.21952185034751892, + "qparams": { + "group_size": 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": 0.1901346892118454, + "qparams": { + "group_size": 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": 0.17812499403953552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15605434775352478, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09997731447219849, + "qparams": { + "group_size": 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": 0.08892438560724258, + "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": 0.12146487087011337, + "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": 0.1110268384218216, + "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": 0.10447622090578079, + "qparams": { + "group_size": 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": 0.08354930579662323, + "qparams": { + "group_size": 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": 0.078175850212574, + "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": 0.06201468035578728, + "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": 0.0534738227725029, + "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": 0.04848712310194969, + "qparams": { + "group_size": 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": 0.0472421832382679, + "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": 0.0312234815210104, + "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": 0.025931773707270622, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025678612291812897, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022798525169491768, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02189745381474495, + "qparams": { + "group_size": 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": 0.017080850899219513, + "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": 0.01759522408246994, + "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": 0.015498876571655273, + "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": 0.012871240265667439, + "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": 0.09997731447219849, + "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": 0.09997731447219849, + "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": 0.10860546678304672, + "qparams": { + "group_size": 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": 0.10234134644269943, + "qparams": { + "group_size": 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": 0.09968496114015579, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09052839130163193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05130011588335037, + "qparams": { + "group_size": 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": 0.048737481236457825, + "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": 0.05814586207270622, + "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": 0.053600963205099106, + "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": 0.051757991313934326, + "qparams": { + "group_size": 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": 0.04597456008195877, + "qparams": { + "group_size": 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": 0.04406115040183067, + "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": 0.029622908681631088, + "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": 0.02570238709449768, + "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": 0.024615563452243805, + "qparams": { + "group_size": 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": 0.024359747767448425, + "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": 0.014812861569225788, + "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": 0.012703240849077702, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012490436434745789, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011619927361607552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011456968262791634, + "qparams": { + "group_size": 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": 0.00772926164790988, + "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": 0.007790617179125547, + "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": 0.007336755748838186, + "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": 0.005179809406399727, + "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": 0.09968496114015579, + "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": 0.09968496114015579, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "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": 0.08902016282081604, + "qparams": { + "group_size": 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": 0.08370526880025864, + "qparams": { + "group_size": 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": 0.08076619356870651, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07309335470199585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04169737547636032, + "qparams": { + "group_size": 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": 0.03927740827202797, + "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": 0.048149511218070984, + "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": 0.04459565505385399, + "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": 0.042099688202142715, + "qparams": { + "group_size": 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": 0.03722735866904259, + "qparams": { + "group_size": 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": 0.03557148575782776, + "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": 0.024265646934509277, + "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": 0.021272653713822365, + "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": 0.019966930150985718, + "qparams": { + "group_size": 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": 0.019660159945487976, + "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": 0.012154781259596348, + "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": 0.010232262313365936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009994125925004482, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.009334435686469078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009123207069933414, + "qparams": { + "group_size": 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": 0.006324701942503452, + "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": 0.006257878616452217, + "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": 0.005921944044530392, + "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": 0.003992103040218353, + "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": 0.08902016282081604, + "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": 0.08902016282081604, + "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": 0.22437132894992828, + "qparams": { + "group_size": 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": 0.21152859926223755, + "qparams": { + "group_size": 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": 0.20682063698768616, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18766887485980988, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10587763041257858, + "qparams": { + "group_size": 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": 0.10094553977251053, + "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": 0.11906085163354874, + "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": 0.10921693593263626, + "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": 0.10676632076501846, + "qparams": { + "group_size": 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": 0.09472104161977768, + "qparams": { + "group_size": 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": 0.09076796472072601, + "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": 0.06071348488330841, + "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": 0.052311088889837265, + "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": 0.05060439929366112, + "qparams": { + "group_size": 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": 0.05020159110426903, + "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": 0.03033282607793808, + "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": 0.02553914301097393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02516048029065132, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02319650538265705, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02292560786008835, + "qparams": { + "group_size": 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": 0.0154775595292449, + "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": 0.01471781451255083, + "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": 0.014730975031852722, + "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": 0.00861942209303379, + "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": 0.09472104161977768, + "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": 0.09472104161977768, + "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": 0.18370285630226135, + "qparams": { + "group_size": 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": 0.16703179478645325, + "qparams": { + "group_size": 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": 0.15978290140628815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1397729218006134, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08500713109970093, + "qparams": { + "group_size": 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": 0.07836312055587769, + "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": 0.10035144537687302, + "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": 0.09217742830514908, + "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": 0.0871877446770668, + "qparams": { + "group_size": 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": 0.0731835663318634, + "qparams": { + "group_size": 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": 0.06851112842559814, + "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": 0.05151093751192093, + "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": 0.044536691159009933, + "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": 0.04127607122063637, + "qparams": { + "group_size": 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": 0.040491022169589996, + "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": 0.025912683457136154, + "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": 0.02209712192416191, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021488990634679794, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.019820941612124443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019318794831633568, + "qparams": { + "group_size": 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": 0.014101354405283928, + "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": 0.014865590259432793, + "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": 0.01307634823024273, + "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": 0.01102143619209528, + "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": 0.10035144537687302, + "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": 0.10035144537687302, + "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": 0.16767600178718567, + "qparams": { + "group_size": 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": 0.1587691307067871, + "qparams": { + "group_size": 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": 0.15566082298755646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14227594435214996, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07923096418380737, + "qparams": { + "group_size": 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": 0.07602301985025406, + "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": 0.08791810274124146, + "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": 0.08135005831718445, + "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": 0.07987305521965027, + "qparams": { + "group_size": 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": 0.07158058881759644, + "qparams": { + "group_size": 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": 0.06869780272245407, + "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": 0.04470643028616905, + "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": 0.03894765302538872, + "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": 0.037995368242263794, + "qparams": { + "group_size": 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": 0.0377705954015255, + "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": 0.022357486188411713, + "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": 0.019636044278740883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01940787211060524, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01809401996433735, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01795022562146187, + "qparams": { + "group_size": 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": 0.011770178563892841, + "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": 0.011987326666712761, + "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": 0.011451087892055511, + "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": 0.008132635615766048, + "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": 0.08791810274124146, + "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": 0.08791810274124146, + "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": 0.20740221440792084, + "qparams": { + "group_size": 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": 0.1963403820991516, + "qparams": { + "group_size": 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": 0.19265897572040558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.17610959708690643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09786760807037354, + "qparams": { + "group_size": 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": 0.093943752348423, + "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": 0.10832799226045609, + "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": 0.10027991235256195, + "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": 0.09862764179706573, + "qparams": { + "group_size": 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": 0.08844519406557083, + "qparams": { + "group_size": 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": 0.08477679640054703, + "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": 0.055024661123752594, + "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": 0.04789275303483009, + "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": 0.04678400978446007, + "qparams": { + "group_size": 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": 0.04652188718318939, + "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": 0.027471287176012993, + "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": 0.023830024525523186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023558078333735466, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02188054844737053, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021710915490984917, + "qparams": { + "group_size": 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": 0.01421440951526165, + "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": 0.014002141542732716, + "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": 0.013834328390657902, + "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": 0.008813964203000069, + "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": 0.10027991235256195, + "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": 0.09862764179706573, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.21512356400489807, + "qparams": { + "group_size": 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": 0.1847703754901886, + "qparams": { + "group_size": 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": 0.17231225967407227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15008069574832916, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09768564999103546, + "qparams": { + "group_size": 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": 0.08634604513645172, + "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": 0.11888499557971954, + "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": 0.10907895863056183, + "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": 0.10242947936058044, + "qparams": { + "group_size": 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": 0.08104345202445984, + "qparams": { + "group_size": 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": 0.07552723586559296, + "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": 0.06092831492424011, + "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": 0.05262541025876999, + "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": 0.04741393029689789, + "qparams": { + "group_size": 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": 0.04611556977033615, + "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": 0.030758440494537354, + "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": 0.02543647028505802, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02517710253596306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02229107916355133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021358570083975792, + "qparams": { + "group_size": 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": 0.016885031014680862, + "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": 0.017408065497875214, + "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": 0.015266693197190762, + "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": 0.012802371755242348, + "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": 0.09768564999103546, + "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": 0.09768564999103546, + "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": 0.11421281099319458, + "qparams": { + "group_size": 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": 0.10783273726701736, + "qparams": { + "group_size": 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": 0.10507911443710327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09572277963161469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.053992122411727905, + "qparams": { + "group_size": 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": 0.05141079053282738, + "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": 0.06102816015481949, + "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": 0.0563330315053463, + "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": 0.05445307865738869, + "qparams": { + "group_size": 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": 0.04852345213294029, + "qparams": { + "group_size": 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": 0.0465632863342762, + "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": 0.031064502894878387, + "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": 0.026974260807037354, + "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": 0.025897188112139702, + "qparams": { + "group_size": 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": 0.02563832327723503, + "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": 0.015533890575170517, + "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": 0.013323481194674969, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.013101726770401001, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.012211030349135399, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012050054036080837, + "qparams": { + "group_size": 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": 0.008082862943410873, + "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": 0.008114323019981384, + "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": 0.007696474436670542, + "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": 0.005331003572791815, + "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": 0.09572277963161469, + "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": 0.09572277963161469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.k_proj", + "numel": 2097152, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1904296875, + "total_bits": 4593664.0, + "err": 0.09104762971401215, + "qparams": { + "group_size": 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": 0.08561031520366669, + "qparams": { + "group_size": 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": 0.08281014114618301, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07519184052944183, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04267677664756775, + "qparams": { + "group_size": 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": 0.040300462394952774, + "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": 0.04906221479177475, + "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": 0.04553768038749695, + "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": 0.04311135783791542, + "qparams": { + "group_size": 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": 0.038302231580019, + "qparams": { + "group_size": 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": 0.03672243282198906, + "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": 0.024868449196219444, + "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": 0.021767647936940193, + "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": 0.020472342148423195, + "qparams": { + "group_size": 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": 0.02015301026403904, + "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": 0.012428950518369675, + "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": 0.010439773090183735, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010195732116699219, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.00953192450106144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009320254437625408, + "qparams": { + "group_size": 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": 0.006429458037018776, + "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": 0.0063161104917526245, + "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": 0.0060160597786307335, + "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": 0.0039023628924041986, + "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": 0.09104762971401215, + "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": 0.09104762971401215, + "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": 0.24163049459457397, + "qparams": { + "group_size": 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": 0.22785496711730957, + "qparams": { + "group_size": 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": 0.2230636179447174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.20270752906799316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.11406895518302917, + "qparams": { + "group_size": 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": 0.10905562341213226, + "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": 0.12785182893276215, + "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": 0.1177053302526474, + "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": 0.11514931172132492, + "qparams": { + "group_size": 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": 0.10235165804624557, + "qparams": { + "group_size": 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": 0.09815073013305664, + "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": 0.06519665569067001, + "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": 0.056237880140542984, + "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": 0.05461212620139122, + "qparams": { + "group_size": 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": 0.05416978523135185, + "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": 0.03251929581165314, + "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": 0.027501605451107025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02711530216038227, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.025015641003847122, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02477947250008583, + "qparams": { + "group_size": 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": 0.016608739271759987, + "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": 0.01579180359840393, + "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": 0.01589152216911316, + "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": 0.009201090782880783, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option_max": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09815073013305664, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + "best_option": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09815073013305664, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "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": 0.16172441840171814, + "qparams": { + "group_size": 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": 0.13552618026733398, + "qparams": { + "group_size": 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": 0.1265438199043274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10780717432498932, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07430293411016464, + "qparams": { + "group_size": 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": 0.06467629224061966, + "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": 0.08763323724269867, + "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": 0.08007025718688965, + "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": 0.07683586329221725, + "qparams": { + "group_size": 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": 0.05803042650222778, + "qparams": { + "group_size": 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": 0.05221680551767349, + "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": 0.04510147497057915, + "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": 0.038911182433366776, + "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": 0.036287110298871994, + "qparams": { + "group_size": 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": 0.03564167395234108, + "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": 0.02313193306326866, + "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": 0.019738715142011642, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019124548882246017, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.016728000715374947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01630678027868271, + "qparams": { + "group_size": 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": 0.013331608846783638, + "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": 0.013606278225779533, + "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": 0.01255291048437357, + "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": 0.0104904780164361, + "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": 0.08763323724269867, + "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": 0.08763323724269867, + "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": 0.16854973137378693, + "qparams": { + "group_size": 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": 0.1596357226371765, + "qparams": { + "group_size": 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": 0.15663591027259827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14324381947517395, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0796174481511116, + "qparams": { + "group_size": 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": 0.07642364501953125, + "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": 0.08819039165973663, + "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": 0.08169078826904297, + "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": 0.08024726808071136, + "qparams": { + "group_size": 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": 0.07197768986225128, + "qparams": { + "group_size": 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": 0.06913453340530396, + "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": 0.04485552757978439, + "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": 0.03907131776213646, + "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": 0.03813346475362778, + "qparams": { + "group_size": 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": 0.037918008863925934, + "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": 0.022398481145501137, + "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": 0.019663486629724503, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019440095871686935, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.018120912835001945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017982471734285355, + "qparams": { + "group_size": 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": 0.011721532791852951, + "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": 0.011926802806556225, + "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": 0.011407016776502132, + "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": 0.008002502843737602, + "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": 0.08819039165973663, + "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": 0.08819039165973663, + "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": 0.20132163166999817, + "qparams": { + "group_size": 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": 0.1907724291086197, + "qparams": { + "group_size": 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": 0.18716435134410858, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.17121101915836334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0949733629822731, + "qparams": { + "group_size": 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": 0.0912070944905281, + "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": 0.10514514893293381, + "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": 0.09729351103305817, + "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": 0.095711849629879, + "qparams": { + "group_size": 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": 0.08587883412837982, + "qparams": { + "group_size": 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": 0.08239845186471939, + "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": 0.05334419012069702, + "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": 0.046423450112342834, + "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": 0.0453806109726429, + "qparams": { + "group_size": 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": 0.04513237625360489, + "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": 0.026617227122187614, + "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": 0.02311922051012516, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022857118397951126, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021239014342427254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02108559012413025, + "qparams": { + "group_size": 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": 0.01375990267843008, + "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": 0.013585038483142853, + "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": 0.01339708175510168, + "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": 0.008547956123948097, + "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": 0.09729351103305817, + "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": 0.09729351103305817, + "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": 0.2132362276315689, + "qparams": { + "group_size": 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": 0.18266019225120544, + "qparams": { + "group_size": 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": 0.17007552087306976, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14795710146427155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09679671376943588, + "qparams": { + "group_size": 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": 0.0852617546916008, + "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": 0.118677519261837, + "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": 0.10826045274734497, + "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": 0.10152960568666458, + "qparams": { + "group_size": 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": 0.08002237975597382, + "qparams": { + "group_size": 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": 0.07442399859428406, + "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": 0.060548193752765656, + "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": 0.05219721421599388, + "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": 0.04697243124246597, + "qparams": { + "group_size": 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": 0.04567607864737511, + "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": 0.03052779659628868, + "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": 0.025195090100169182, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024941764771938324, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022013112902641296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021087629720568657, + "qparams": { + "group_size": 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": 0.01675487868487835, + "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": 0.017233705148100853, + "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": 0.015100592747330666, + "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": 0.012669658288359642, + "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": 0.09679671376943588, + "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": 0.09679671376943588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.11051224172115326, + "qparams": { + "group_size": 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": 0.10419156402349472, + "qparams": { + "group_size": 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": 0.10155877470970154, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09244909137487411, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05218307301402092, + "qparams": { + "group_size": 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": 0.04969630017876625, + "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": 0.059029996395111084, + "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": 0.05438973754644394, + "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": 0.05264351889491081, + "qparams": { + "group_size": 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": 0.04685221239924431, + "qparams": { + "group_size": 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": 0.04492190107703209, + "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": 0.030047014355659485, + "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": 0.026064425706863403, + "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": 0.025033356621861458, + "qparams": { + "group_size": 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": 0.02478518709540367, + "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": 0.01502399705350399, + "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": 0.012891699559986591, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012682544067502022, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011807386763393879, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011654006317257881, + "qparams": { + "group_size": 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": 0.00783581007272005, + "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": 0.007866530679166317, + "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": 0.007462434470653534, + "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": 0.005205844528973103, + "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": 0.09244909137487411, + "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": 0.09244909137487411, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.k_proj", + "numel": 2097152, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1904296875, + "total_bits": 4593664.0, + "err": 0.08971717953681946, + "qparams": { + "group_size": 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": 0.08433231711387634, + "qparams": { + "group_size": 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": 0.08150684833526611, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07401770353317261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.042146265506744385, + "qparams": { + "group_size": 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": 0.039599817246198654, + "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": 0.04858468845486641, + "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": 0.044934920966625214, + "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": 0.04248582571744919, + "qparams": { + "group_size": 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": 0.03766784816980362, + "qparams": { + "group_size": 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": 0.03616722300648689, + "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": 0.024647129699587822, + "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": 0.021585257723927498, + "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": 0.02016008086502552, + "qparams": { + "group_size": 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": 0.019836945459246635, + "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": 0.012284801341593266, + "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": 0.01028490997850895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010031279176473618, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.009369024075567722, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0091466773301363, + "qparams": { + "group_size": 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": 0.006368021946400404, + "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": 0.006228155922144651, + "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": 0.005940990522503853, + "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": 0.003859166754409671, + "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": 0.08971717953681946, + "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": 0.08971717953681946, + "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": 0.22591474652290344, + "qparams": { + "group_size": 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": 0.2122095227241516, + "qparams": { + "group_size": 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": 0.20760944485664368, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18855378031730652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10633613169193268, + "qparams": { + "group_size": 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": 0.10136768221855164, + "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": 0.11957936733961105, + "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": 0.11013539135456085, + "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": 0.10736806690692902, + "qparams": { + "group_size": 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": 0.09522905200719833, + "qparams": { + "group_size": 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": 0.09113582968711853, + "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": 0.0609050914645195, + "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": 0.052711281925439835, + "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": 0.050903867930173874, + "qparams": { + "group_size": 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": 0.05045906454324722, + "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": 0.030488017946481705, + "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": 0.02572481520473957, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025344526395201683, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.023392850533127785, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023131858557462692, + "qparams": { + "group_size": 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": 0.015575719065964222, + "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": 0.01496864203363657, + "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": 0.014828318729996681, + "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": 0.008941921405494213, + "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": 0.09522905200719833, + "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": 0.09522905200719833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "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": 0.21215452253818512, + "qparams": { + "group_size": 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": 0.1878107488155365, + "qparams": { + "group_size": 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": 0.17976853251457214, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15079960227012634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09839324653148651, + "qparams": { + "group_size": 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": 0.08983959257602692, + "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": 0.11315902322530746, + "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": 0.10346149653196335, + "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": 0.10029904544353485, + "qparams": { + "group_size": 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": 0.07943683117628098, + "qparams": { + "group_size": 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": 0.07290274649858475, + "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": 0.058049120008945465, + "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": 0.04992888122797012, + "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": 0.04760118946433067, + "qparams": { + "group_size": 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": 0.04707983881235123, + "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": 0.02921219915151596, + "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": 0.025298694148659706, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024738825857639313, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02164924331009388, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021285902708768845, + "qparams": { + "group_size": 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": 0.01592913642525673, + "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": 0.016575545072555542, + "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": 0.015207821503281593, + "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": 0.012230747379362583, + "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": 3.1878662109375, + "total_bits": 53483520.0, + "err": 0.10029904544353485, + "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": 3.2191162109375, + "total_bits": 54007808.0, + "err": 0.09839324653148651, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.1672891229391098, + "qparams": { + "group_size": 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": 0.15846669673919678, + "qparams": { + "group_size": 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": 0.15543213486671448, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14226225018501282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0790238231420517, + "qparams": { + "group_size": 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": 0.07586358487606049, + "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": 0.08758985996246338, + "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": 0.08110526204109192, + "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": 0.07963989675045013, + "qparams": { + "group_size": 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": 0.07152018696069717, + "qparams": { + "group_size": 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": 0.06861460208892822, + "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": 0.04455285146832466, + "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": 0.03883231803774834, + "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": 0.03789343684911728, + "qparams": { + "group_size": 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": 0.03766680881381035, + "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": 0.022279057651758194, + "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": 0.019564781337976456, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019342636689543724, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.018049538135528564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0179069172590971, + "qparams": { + "group_size": 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": 0.01172059215605259, + "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": 0.01192646473646164, + "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": 0.011408657766878605, + "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": 0.008066645823419094, + "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": 0.08758985996246338, + "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": 0.08758985996246338, + "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": 0.19872945547103882, + "qparams": { + "group_size": 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": 0.1882607638835907, + "qparams": { + "group_size": 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": 0.1847413033246994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16900403797626495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09368875622749329, + "qparams": { + "group_size": 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": 0.0899675041437149, + "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": 0.1037307158112526, + "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": 0.09599647670984268, + "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": 0.09441637992858887, + "qparams": { + "group_size": 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": 0.08474663645029068, + "qparams": { + "group_size": 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": 0.0813521072268486, + "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": 0.05265282094478607, + "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": 0.0458073653280735, + "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": 0.0447591133415699, + "qparams": { + "group_size": 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": 0.044511161744594574, + "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": 0.02626882866024971, + "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": 0.022793162614107132, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022533565759658813, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020951882004737854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020797137171030045, + "qparams": { + "group_size": 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": 0.013566385954618454, + "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": 0.013380912132561207, + "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": 0.013202005997300148, + "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": 0.008406980894505978, + "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": 0.09599647670984268, + "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": 0.09599647670984268, + "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": 0.21122802793979645, + "qparams": { + "group_size": 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": 0.18011564016342163, + "qparams": { + "group_size": 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": 0.1669766902923584, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14456190168857574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09575597941875458, + "qparams": { + "group_size": 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": 0.0837896466255188, + "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": 0.11827871948480606, + "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": 0.1078585833311081, + "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": 0.10063378512859344, + "qparams": { + "group_size": 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": 0.07874087989330292, + "qparams": { + "group_size": 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": 0.07317784428596497, + "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": 0.06078845262527466, + "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": 0.05209872126579285, + "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": 0.04651391878724098, + "qparams": { + "group_size": 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": 0.04513033106923103, + "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": 0.030795715749263763, + "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": 0.02500961162149906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024712996557354927, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021802542731165886, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020801346749067307, + "qparams": { + "group_size": 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": 0.01712976023554802, + "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": 0.01724942959845066, + "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": 0.015351377427577972, + "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": 0.012706412933766842, + "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": 0.10063378512859344, + "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": 0.10063378512859344, + "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": 0.11091136932373047, + "qparams": { + "group_size": 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": 0.10437972098588943, + "qparams": { + "group_size": 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": 0.10154323279857635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09236638993024826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05237036570906639, + "qparams": { + "group_size": 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": 0.04969888925552368, + "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": 0.05962953343987465, + "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": 0.05491239205002785, + "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": 0.052843742072582245, + "qparams": { + "group_size": 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": 0.04690813273191452, + "qparams": { + "group_size": 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": 0.04499626159667969, + "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": 0.03030390478670597, + "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": 0.026285771280527115, + "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": 0.025127435103058815, + "qparams": { + "group_size": 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": 0.02485317550599575, + "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": 0.0151491304859519, + "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": 0.012936730869114399, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01270824484527111, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011825796216726303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011648681946098804, + "qparams": { + "group_size": 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": 0.007908489555120468, + "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": 0.007911822758615017, + "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": 0.007499916944652796, + "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": 0.005202317610383034, + "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": 0.09236638993024826, + "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": 0.09236638993024826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.k_proj", + "numel": 2097152, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1904296875, + "total_bits": 4593664.0, + "err": 0.08738277107477188, + "qparams": { + "group_size": 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": 0.08181724697351456, + "qparams": { + "group_size": 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": 0.07880271226167679, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07146298140287399, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04091254994273186, + "qparams": { + "group_size": 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": 0.03830723091959953, + "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": 0.04751225560903549, + "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": 0.04418620839715004, + "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": 0.041331302374601364, + "qparams": { + "group_size": 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": 0.03657573461532593, + "qparams": { + "group_size": 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": 0.03501804545521736, + "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": 0.024054791778326035, + "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": 0.02112669125199318, + "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": 0.019590919837355614, + "qparams": { + "group_size": 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": 0.01923055201768875, + "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": 0.012026694603264332, + "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": 0.01001126877963543, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009736096486449242, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.009100147522985935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00886848010122776, + "qparams": { + "group_size": 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": 0.006242634728550911, + "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": 0.006130333058536053, + "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": 0.005771649535745382, + "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": 0.003809733549132943, + "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": 0.08738277107477188, + "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": 0.08738277107477188, + "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": 0.21373368799686432, + "qparams": { + "group_size": 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": 0.19922670722007751, + "qparams": { + "group_size": 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": 0.19344720244407654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.17444643378257751, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10023066401481628, + "qparams": { + "group_size": 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": 0.09450603276491165, + "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": 0.11413894593715668, + "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": 0.10487271100282669, + "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": 0.10131807625293732, + "qparams": { + "group_size": 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": 0.08858295530080795, + "qparams": { + "group_size": 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": 0.08478029817342758, + "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": 0.05814776569604874, + "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": 0.050181470811367035, + "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": 0.04791441559791565, + "qparams": { + "group_size": 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": 0.04738098382949829, + "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": 0.029076524078845978, + "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": 0.024207036942243576, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02371489629149437, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021771283820271492, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02142276242375374, + "qparams": { + "group_size": 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": 0.014850097708404064, + "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": 0.014155296608805656, + "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": 0.013935253024101257, + "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": 0.00833459198474884, + "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": 0.10023066401481628, + "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": 0.10023066401481628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.1490703970193863, + "qparams": { + "group_size": 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": 0.13541434705257416, + "qparams": { + "group_size": 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": 0.13109371066093445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.11014781147241592, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07005779445171356, + "qparams": { + "group_size": 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": 0.06518702954053879, + "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": 0.07903552800416946, + "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": 0.0723753497004509, + "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": 0.07099324464797974, + "qparams": { + "group_size": 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": 0.057184673845767975, + "qparams": { + "group_size": 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": 0.0523974783718586, + "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": 0.040687285363674164, + "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": 0.034960389137268066, + "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": 0.03390447795391083, + "qparams": { + "group_size": 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": 0.03366551920771599, + "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": 0.020388035103678703, + "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": 0.017967842519283295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01767442189157009, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.015457240864634514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.015294784680008888, + "qparams": { + "group_size": 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": 0.01102755218744278, + "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": 0.01162286289036274, + "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": 0.010695718228816986, + "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": 0.008597173728048801, + "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": 0.07903552800416946, + "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": 0.07903552800416946, + "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": 0.17214921116828918, + "qparams": { + "group_size": 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": 0.16311047971248627, + "qparams": { + "group_size": 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": 0.16010993719100952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14662975072860718, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08130050450563431, + "qparams": { + "group_size": 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": 0.07806894183158875, + "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": 0.08990795910358429, + "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": 0.08331450819969177, + "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": 0.08191321790218353, + "qparams": { + "group_size": 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": 0.07356380671262741, + "qparams": { + "group_size": 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": 0.07065577059984207, + "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": 0.04573221132159233, + "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": 0.03983544558286667, + "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": 0.0389067679643631, + "qparams": { + "group_size": 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": 0.03868750482797623, + "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": 0.022837404161691666, + "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": 0.019979054108262062, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019753776490688324, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.018404988572001457, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.018266959115862846, + "qparams": { + "group_size": 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": 0.01191961020231247, + "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": 0.011987406760454178, + "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": 0.011611764319241047, + "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": 0.007890675216913223, + "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": 0.08990795910358429, + "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": 0.08990795910358429, + "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": 0.19788822531700134, + "qparams": { + "group_size": 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": 0.18760399520397186, + "qparams": { + "group_size": 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": 0.18414434790611267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16861920058727264, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09329470992088318, + "qparams": { + "group_size": 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": 0.08968546986579895, + "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": 0.10324548184871674, + "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": 0.0955430194735527, + "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": 0.09402521699666977, + "qparams": { + "group_size": 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": 0.08450200408697128, + "qparams": { + "group_size": 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": 0.08110283315181732, + "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": 0.0523744560778141, + "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": 0.045602425932884216, + "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": 0.044591035693883896, + "qparams": { + "group_size": 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": 0.04435083642601967, + "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": 0.02614276111125946, + "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": 0.022731272503733635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022481108084321022, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020914005115628242, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020758705213665962, + "qparams": { + "group_size": 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": 0.01352263055741787, + "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": 0.013369366526603699, + "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": 0.01317504607141018, + "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": 0.008453311398625374, + "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": 0.0955430194735527, + "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": 0.0955430194735527, + "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": 0.21092897653579712, + "qparams": { + "group_size": 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": 0.1786613166332245, + "qparams": { + "group_size": 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": 0.16524939239025116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1424657106399536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0954306498169899, + "qparams": { + "group_size": 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": 0.08324271440505981, + "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": 0.11757701635360718, + "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": 0.10751228779554367, + "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": 0.10045178234577179, + "qparams": { + "group_size": 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": 0.07783680409193039, + "qparams": { + "group_size": 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": 0.07216804474592209, + "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": 0.060304634273052216, + "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": 0.051942791789770126, + "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": 0.04647213593125343, + "qparams": { + "group_size": 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": 0.04511808604001999, + "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": 0.030516577884554863, + "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": 0.025155242532491684, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024886613711714745, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021863505244255066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020872365683317184, + "qparams": { + "group_size": 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": 0.016981760039925575, + "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": 0.01751052215695381, + "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": 0.015255918726325035, + "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": 0.013090068474411964, + "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": 0.10045178234577179, + "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": 0.10045178234577179, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.10834978520870209, + "qparams": { + "group_size": 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": 0.10206568241119385, + "qparams": { + "group_size": 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": 0.0993383526802063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.090414859354496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.051185425370931625, + "qparams": { + "group_size": 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": 0.04862246289849281, + "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": 0.05818446725606918, + "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": 0.05361935496330261, + "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": 0.05163934454321861, + "qparams": { + "group_size": 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": 0.045909322798252106, + "qparams": { + "group_size": 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": 0.04405329003930092, + "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": 0.0296026524156332, + "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": 0.025673074647784233, + "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": 0.024543065577745438, + "qparams": { + "group_size": 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": 0.024279436096549034, + "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": 0.014789631590247154, + "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": 0.012610072270035744, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012382063083350658, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011534862220287323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011361295357346535, + "qparams": { + "group_size": 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": 0.0077077229507267475, + "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": 0.0076753743924200535, + "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": 0.00731310760602355, + "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": 0.00500553660094738, + "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": 0.0993383526802063, + "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": 0.0993383526802063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "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": 0.08714371919631958, + "qparams": { + "group_size": 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": 0.08166541904211044, + "qparams": { + "group_size": 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": 0.07849512994289398, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07111195474863052, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0407814159989357, + "qparams": { + "group_size": 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": 0.03822038322687149, + "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": 0.04752695560455322, + "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": 0.04410196468234062, + "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": 0.0412101149559021, + "qparams": { + "group_size": 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": 0.03637279197573662, + "qparams": { + "group_size": 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": 0.03484727814793587, + "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": 0.02402416802942753, + "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": 0.021031612530350685, + "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": 0.019520943984389305, + "qparams": { + "group_size": 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": 0.019170546904206276, + "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": 0.011994306929409504, + "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": 0.009973259642720222, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009702006354928017, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.009050773456692696, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008821587078273296, + "qparams": { + "group_size": 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": 0.0062375543639063835, + "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": 0.00607881136238575, + "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": 0.005772814620286226, + "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": 0.003756885416805744, + "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": 0.08714371919631958, + "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": 0.08714371919631958, + "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": 0.21667760610580444, + "qparams": { + "group_size": 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": 0.2026033252477646, + "qparams": { + "group_size": 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": 0.1973373293876648, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1787959635257721, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10166134685277939, + "qparams": { + "group_size": 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": 0.09634257853031158, + "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": 0.11529745161533356, + "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": 0.10605431348085403, + "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": 0.10271622240543365, + "qparams": { + "group_size": 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": 0.09045077115297318, + "qparams": { + "group_size": 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": 0.08666107058525085, + "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": 0.05879422277212143, + "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": 0.05061472952365875, + "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": 0.04856632649898529, + "qparams": { + "group_size": 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": 0.048036959022283554, + "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": 0.029333150014281273, + "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": 0.024581894278526306, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02413300797343254, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022247999906539917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021915307268500328, + "qparams": { + "group_size": 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": 0.01500589307397604, + "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": 0.0143976341933012, + "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": 0.014153817668557167, + "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": 0.008553491905331612, + "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": 0.09045077115297318, + "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": 0.09045077115297318, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "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": 0.2135726511478424, + "qparams": { + "group_size": 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": 0.18934477865695953, + "qparams": { + "group_size": 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": 0.18100422620773315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14828255772590637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09948687255382538, + "qparams": { + "group_size": 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": 0.09089731425046921, + "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": 0.11442334949970245, + "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": 0.1048336923122406, + "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": 0.10142021626234055, + "qparams": { + "group_size": 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": 0.07889220118522644, + "qparams": { + "group_size": 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": 0.07170448452234268, + "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": 0.05888547748327255, + "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": 0.05080870911478996, + "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": 0.04840860143303871, + "qparams": { + "group_size": 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": 0.04782639071345329, + "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": 0.029631346464157104, + "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": 0.026225103065371513, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02568298950791359, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022396210581064224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02202209271490574, + "qparams": { + "group_size": 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": 0.01643957756459713, + "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": 0.01778748817741871, + "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": 0.01571059226989746, + "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": 0.013747751712799072, + "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": 3.2191162109375, + "total_bits": 54007808.0, + "err": 0.09948687255382538, + "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": 3.2191162109375, + "total_bits": 54007808.0, + "err": 0.09948687255382538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.17798514664173126, + "qparams": { + "group_size": 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": 0.16871699690818787, + "qparams": { + "group_size": 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": 0.16559620201587677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15161943435668945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08407538384199142, + "qparams": { + "group_size": 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": 0.08076492697000504, + "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": 0.0930403620004654, + "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": 0.08619898557662964, + "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": 0.08473807573318481, + "qparams": { + "group_size": 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": 0.07608132064342499, + "qparams": { + "group_size": 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": 0.0730394572019577, + "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": 0.04730026796460152, + "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": 0.04118200019001961, + "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": 0.04022480919957161, + "qparams": { + "group_size": 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": 0.039996933192014694, + "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": 0.023611249402165413, + "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": 0.020595362409949303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02036128006875515, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.018958805128932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01881251484155655, + "qparams": { + "group_size": 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": 0.012297228910028934, + "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": 0.012269724160432816, + "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": 0.011980971321463585, + "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": 0.00795080978423357, + "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": 0.0930403620004654, + "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": 0.0930403620004654, + "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": 0.20219630002975464, + "qparams": { + "group_size": 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": 0.19166894257068634, + "qparams": { + "group_size": 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": 0.18817099928855896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.17234162986278534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09540951997041702, + "qparams": { + "group_size": 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": 0.09168222546577454, + "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": 0.10537886619567871, + "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": 0.09763936698436737, + "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": 0.09613720327615738, + "qparams": { + "group_size": 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": 0.0863567441701889, + "qparams": { + "group_size": 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": 0.0828857496380806, + "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": 0.053533926606178284, + "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": 0.046616651117801666, + "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": 0.045592546463012695, + "qparams": { + "group_size": 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": 0.0453558973968029, + "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": 0.026710830628871918, + "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": 0.023209672421216965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02294846624135971, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02133501134812832, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021180620416998863, + "qparams": { + "group_size": 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": 0.0138021819293499, + "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": 0.013601602986454964, + "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": 0.013451370410621166, + "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": 0.008532841689884663, + "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": 0.09763936698436737, + "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": 0.09763936698436737, + "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": 0.20662231743335724, + "qparams": { + "group_size": 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": 0.17405842244625092, + "qparams": { + "group_size": 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": 0.16050146520137787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.13801805675029755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09326662868261337, + "qparams": { + "group_size": 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": 0.0808754712343216, + "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": 0.1158050000667572, + "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": 0.10536126047372818, + "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": 0.0983600988984108, + "qparams": { + "group_size": 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": 0.07557310909032822, + "qparams": { + "group_size": 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": 0.07006604224443436, + "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": 0.058969803154468536, + "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": 0.05074836313724518, + "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": 0.04526892676949501, + "qparams": { + "group_size": 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": 0.04389927163720131, + "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": 0.029696861281991005, + "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": 0.02424989454448223, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023970916867256165, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020891282707452774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019880173727869987, + "qparams": { + "group_size": 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": 0.016273053362965584, + "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": 0.01662830449640751, + "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": 0.01452731341123581, + "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": 0.012104717083275318, + "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": 0.0983600988984108, + "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": 0.0983600988984108, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.10484173893928528, + "qparams": { + "group_size": 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": 0.09836508333683014, + "qparams": { + "group_size": 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": 0.09549044072628021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08663872629404068, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04936742037534714, + "qparams": { + "group_size": 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": 0.04669977352023125, + "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": 0.05642634630203247, + "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": 0.05198314040899277, + "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": 0.049840278923511505, + "qparams": { + "group_size": 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": 0.04409297555685043, + "qparams": { + "group_size": 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": 0.04222838953137398, + "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": 0.028670983389019966, + "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": 0.024856530129909515, + "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": 0.0236777663230896, + "qparams": { + "group_size": 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": 0.023388946428894997, + "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": 0.01433577761054039, + "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": 0.012161030434072018, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011921080760657787, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011077183298766613, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010894957929849625, + "qparams": { + "group_size": 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": 0.007483516354113817, + "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": 0.007408733479678631, + "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": 0.00707163754850626, + "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": 0.004800054710358381, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09836508333683014, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09836508333683014, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "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": 0.08826477080583572, + "qparams": { + "group_size": 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": 0.08247067779302597, + "qparams": { + "group_size": 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": 0.07918249070644379, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07180192321538925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.041368041187524796, + "qparams": { + "group_size": 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": 0.038604892790317535, + "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": 0.04822839796543121, + "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": 0.04474792629480362, + "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": 0.041795555502176285, + "qparams": { + "group_size": 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": 0.03674744442105293, + "qparams": { + "group_size": 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": 0.03521547466516495, + "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": 0.024502279236912727, + "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": 0.021414171904325485, + "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": 0.01979122869670391, + "qparams": { + "group_size": 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": 0.019406434148550034, + "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": 0.012263390235602856, + "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": 0.010156895034015179, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009856614284217358, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.009224478155374527, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.008975498378276825, + "qparams": { + "group_size": 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": 0.006422619801014662, + "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": 0.006295336410403252, + "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": 0.005916443187743425, + "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": 0.003984873183071613, + "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": 0.08826477080583572, + "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": 0.08826477080583572, + "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": 0.22686181962490082, + "qparams": { + "group_size": 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": 0.2110128402709961, + "qparams": { + "group_size": 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": 0.20515011250972748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18486197292804718, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10628335177898407, + "qparams": { + "group_size": 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": 0.1000789925456047, + "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": 0.12063713371753693, + "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": 0.11105003207921982, + "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": 0.1074128970503807, + "qparams": { + "group_size": 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": 0.09380194544792175, + "qparams": { + "group_size": 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": 0.08949285000562668, + "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": 0.06157540902495384, + "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": 0.05314528942108154, + "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": 0.05077971890568733, + "qparams": { + "group_size": 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": 0.05021771416068077, + "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": 0.030776415020227432, + "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": 0.025699682533740997, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025181768462061882, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.023083023726940155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0227036289870739, + "qparams": { + "group_size": 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": 0.015727562829852104, + "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": 0.015059181489050388, + "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": 0.014768267050385475, + "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": 0.008936882019042969, + "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": 0.09380194544792175, + "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": 0.09380194544792175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "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": 0.20266439020633698, + "qparams": { + "group_size": 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": 0.1587395966053009, + "qparams": { + "group_size": 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": 0.14352810382843018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12779228389263153, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0929255336523056, + "qparams": { + "group_size": 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": 0.07649566978216171, + "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": 0.11076472699642181, + "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": 0.10133674740791321, + "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": 0.09589618444442749, + "qparams": { + "group_size": 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": 0.06837669759988785, + "qparams": { + "group_size": 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": 0.06470556557178497, + "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": 0.05657540261745453, + "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": 0.04976499453186989, + "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": 0.04602500796318054, + "qparams": { + "group_size": 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": 0.04510529339313507, + "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": 0.028865370899438858, + "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": 0.02606985904276371, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.025129983201622963, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02184119075536728, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021285243332386017, + "qparams": { + "group_size": 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": 0.01668040081858635, + "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": 0.019109472632408142, + "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": 0.015625834465026855, + "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": 0.015711482614278793, + "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": 3.1878662109375, + "total_bits": 53483520.0, + "err": 0.09589618444442749, + "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": 3.1878662109375, + "total_bits": 53483520.0, + "err": 0.09589618444442749, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.18207767605781555, + "qparams": { + "group_size": 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": 0.1724930852651596, + "qparams": { + "group_size": 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": 0.1692594438791275, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.154917374253273, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08604906499385834, + "qparams": { + "group_size": 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": 0.0826263353228569, + "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": 0.09522116929292679, + "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": 0.088198222219944, + "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": 0.08673229813575745, + "qparams": { + "group_size": 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": 0.0778060331940651, + "qparams": { + "group_size": 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": 0.07469778507947922, + "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": 0.04844784364104271, + "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": 0.04214427247643471, + "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": 0.041174937039613724, + "qparams": { + "group_size": 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": 0.04094584658741951, + "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": 0.02418794110417366, + "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": 0.021108299493789673, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020871426910161972, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.019426729530096054, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01928229071199894, + "qparams": { + "group_size": 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": 0.012636328116059303, + "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": 0.012623895891010761, + "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": 0.012312976643443108, + "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": 0.008243320509791374, + "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": 0.09522116929292679, + "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": 0.09522116929292679, + "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": 0.2076430320739746, + "qparams": { + "group_size": 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": 0.196807861328125, + "qparams": { + "group_size": 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": 0.1932385265827179, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.17683222889900208, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09799036383628845, + "qparams": { + "group_size": 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": 0.0941586047410965, + "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": 0.10829591006040573, + "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": 0.10030287504196167, + "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": 0.09877616912126541, + "qparams": { + "group_size": 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": 0.08865086734294891, + "qparams": { + "group_size": 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": 0.08507965505123138, + "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": 0.05499764904379845, + "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": 0.04788028448820114, + "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": 0.04682909697294235, + "qparams": { + "group_size": 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": 0.046568743884563446, + "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": 0.02743804268538952, + "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": 0.023798052221536636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02353351190686226, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02185356803238392, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02169625088572502, + "qparams": { + "group_size": 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": 0.01414228044450283, + "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": 0.013889715075492859, + "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": 0.013782358728349209, + "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": 0.008627335540950298, + "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": 0.10030287504196167, + "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": 0.09799036383628845, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.20961083471775055, + "qparams": { + "group_size": 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": 0.17664854228496552, + "qparams": { + "group_size": 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": 0.16279193758964539, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14002983272075653, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09436724334955215, + "qparams": { + "group_size": 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": 0.08183182775974274, + "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": 0.11720333993434906, + "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": 0.10694229602813721, + "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": 0.09959851950407028, + "qparams": { + "group_size": 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": 0.07657662034034729, + "qparams": { + "group_size": 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": 0.07104132324457169, + "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": 0.05968558043241501, + "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": 0.05136711895465851, + "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": 0.0457000806927681, + "qparams": { + "group_size": 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": 0.04427866265177727, + "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": 0.030215444043278694, + "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": 0.024310214444994926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024014852941036224, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020905736833810806, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01983938366174698, + "qparams": { + "group_size": 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": 0.016716403886675835, + "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": 0.016507066786289215, + "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": 0.014937620609998703, + "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": 0.011768876574933529, + "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": 0.09959851950407028, + "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": 0.09959851950407028, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.10630156844854355, + "qparams": { + "group_size": 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": 0.100123830139637, + "qparams": { + "group_size": 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": 0.09749070554971695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08875522762537003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05014466866850853, + "qparams": { + "group_size": 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": 0.04764828085899353, + "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": 0.05693218857049942, + "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": 0.05251370742917061, + "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": 0.050592515617609024, + "qparams": { + "group_size": 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": 0.044969432055950165, + "qparams": { + "group_size": 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": 0.043172437697649, + "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": 0.028884675353765488, + "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": 0.025072254240512848, + "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": 0.02402891032397747, + "qparams": { + "group_size": 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": 0.023771366104483604, + "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": 0.014431829564273357, + "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": 0.012316309846937656, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012100453488528728, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011262350715696812, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011100733652710915, + "qparams": { + "group_size": 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": 0.007512305863201618, + "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": 0.007446106523275375, + "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": 0.00714902114123106, + "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": 0.004794265143573284, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.100123830139637, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.100123830139637, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "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": 0.09438810497522354, + "qparams": { + "group_size": 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": 0.08873894065618515, + "qparams": { + "group_size": 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": 0.08568771183490753, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07781710475683212, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04441934823989868, + "qparams": { + "group_size": 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": 0.0417698472738266, + "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": 0.051328886300325394, + "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": 0.04742772504687309, + "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": 0.04482327401638031, + "qparams": { + "group_size": 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": 0.03968680649995804, + "qparams": { + "group_size": 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": 0.03803405910730362, + "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": 0.02600232884287834, + "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": 0.022642511874437332, + "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": 0.021272534504532814, + "qparams": { + "group_size": 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": 0.020929496735334396, + "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": 0.01299586333334446, + "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": 0.01084850076586008, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010584683157503605, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.009883221238851547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00966706220060587, + "qparams": { + "group_size": 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": 0.006745718419551849, + "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": 0.006576663348823786, + "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": 0.006314126309007406, + "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": 0.004080514889210463, + "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": 0.09438810497522354, + "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": 0.09438810497522354, + "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": 0.22886714339256287, + "qparams": { + "group_size": 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": 0.2145562618970871, + "qparams": { + "group_size": 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": 0.20945851504802704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1897328794002533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10759761929512024, + "qparams": { + "group_size": 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": 0.10223886370658875, + "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": 0.12089258432388306, + "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": 0.11126097291707993, + "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": 0.10866671055555344, + "qparams": { + "group_size": 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": 0.09579378366470337, + "qparams": { + "group_size": 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": 0.09173153340816498, + "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": 0.061544496566057205, + "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": 0.053149282932281494, + "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": 0.0514056496322155, + "qparams": { + "group_size": 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": 0.05098789185285568, + "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": 0.030774343758821487, + "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": 0.026036445051431656, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02562146820127964, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02357218787074089, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023273199796676636, + "qparams": { + "group_size": 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": 0.015777135267853737, + "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": 0.015171052888035774, + "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": 0.01502072811126709, + "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": 0.009120544418692589, + "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": 0.09579378366470337, + "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": 0.09579378366470337, + "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": 0.19206476211547852, + "qparams": { + "group_size": 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": 0.1646968573331833, + "qparams": { + "group_size": 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": 0.15486721694469452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12914176285266876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0887751430273056, + "qparams": { + "group_size": 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": 0.07813674956560135, + "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": 0.10352765023708344, + "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": 0.0954165905714035, + "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": 0.09112543612718582, + "qparams": { + "group_size": 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": 0.06918955594301224, + "qparams": { + "group_size": 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": 0.06351683288812637, + "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": 0.053213126957416534, + "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": 0.04629596695303917, + "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": 0.043295346200466156, + "qparams": { + "group_size": 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": 0.042570166289806366, + "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": 0.026790928095579147, + "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": 0.023584986105561256, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02287382259964943, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.019972972571849823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01950572058558464, + "qparams": { + "group_size": 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": 0.014822195284068584, + "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": 0.016269398853182793, + "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": 0.013910171575844288, + "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": 0.012601540423929691, + "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": 0.0954165905714035, + "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": 0.0954165905714035, + "qparams": { + "group_size": 32, + "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": 0.18371038138866425, + "qparams": { + "group_size": 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": 0.1741011142730713, + "qparams": { + "group_size": 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": 0.17077958583831787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15633627772331238, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0868891179561615, + "qparams": { + "group_size": 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": 0.08342733234167099, + "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": 0.09620577096939087, + "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": 0.08905649185180664, + "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": 0.08758927881717682, + "qparams": { + "group_size": 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": 0.07857359200716019, + "qparams": { + "group_size": 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": 0.0754041075706482, + "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": 0.048989858478307724, + "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": 0.04259197786450386, + "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": 0.04159911349415779, + "qparams": { + "group_size": 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": 0.041379887610673904, + "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": 0.024459317326545715, + "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": 0.02137715369462967, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021138226613402367, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01967604272067547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019525233656167984, + "qparams": { + "group_size": 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": 0.012820356525480747, + "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": 0.012842218391597271, + "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": 0.012497481890022755, + "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": 0.008469905704259872, + "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": 0.09620577096939087, + "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": 0.09620577096939087, + "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": 0.20810922980308533, + "qparams": { + "group_size": 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": 0.19718125462532043, + "qparams": { + "group_size": 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": 0.1936047375202179, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.17718574404716492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09825611114501953, + "qparams": { + "group_size": 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": 0.09438387304544449, + "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": 0.1085214763879776, + "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": 0.10052931308746338, + "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": 0.0990198478102684, + "qparams": { + "group_size": 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": 0.08884686231613159, + "qparams": { + "group_size": 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": 0.08532598614692688, + "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": 0.05514850839972496, + "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": 0.04799380525946617, + "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": 0.04694996774196625, + "qparams": { + "group_size": 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": 0.04669128358364105, + "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": 0.027525337412953377, + "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": 0.023866167291998863, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023608161136507988, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021923447027802467, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021760720759630203, + "qparams": { + "group_size": 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": 0.014227086678147316, + "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": 0.013949678279459476, + "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": 0.013871981762349606, + "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": 0.008690642192959785, + "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": 0.10052931308746338, + "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": 0.09825611114501953, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.2067251354455948, + "qparams": { + "group_size": 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": 0.17356643080711365, + "qparams": { + "group_size": 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": 0.15869227051734924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1366448849439621, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09289238601922989, + "qparams": { + "group_size": 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": 0.079841747879982, + "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": 0.11710997670888901, + "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": 0.10700386017560959, + "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": 0.09823954850435257, + "qparams": { + "group_size": 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": 0.07545007020235062, + "qparams": { + "group_size": 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": 0.07000938802957535, + "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": 0.05934701859951019, + "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": 0.05147223174571991, + "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": 0.045103125274181366, + "qparams": { + "group_size": 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": 0.04348894953727722, + "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": 0.029839055612683296, + "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": 0.024135716259479523, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023744886741042137, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02081497199833393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01964975707232952, + "qparams": { + "group_size": 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": 0.0161274541169405, + "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": 0.016663528978824615, + "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": 0.014119287021458149, + "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": 0.011995567940175533, + "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": 0.09823954850435257, + "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": 0.09823954850435257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.11626093834638596, + "qparams": { + "group_size": 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": 0.10899903625249863, + "qparams": { + "group_size": 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": 0.1058465838432312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09603773057460785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05485115945339203, + "qparams": { + "group_size": 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": 0.05184677243232727, + "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": 0.06251154094934464, + "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": 0.05773104354739189, + "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": 0.05540629103779793, + "qparams": { + "group_size": 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": 0.04890362545847893, + "qparams": { + "group_size": 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": 0.04686461389064789, + "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": 0.031788360327482224, + "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": 0.02762514166533947, + "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": 0.026298711076378822, + "qparams": { + "group_size": 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": 0.025980694219470024, + "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": 0.015886614099144936, + "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": 0.01348006073385477, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.013216483406722546, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01225992664694786, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01205778494477272, + "qparams": { + "group_size": 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": 0.008248567581176758, + "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": 0.008179968222975731, + "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": 0.00779787078499794, + "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": 0.005264753941446543, + "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": 0.09603773057460785, + "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": 0.09603773057460785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.k_proj", + "numel": 2097152, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1904296875, + "total_bits": 4593664.0, + "err": 0.09939263015985489, + "qparams": { + "group_size": 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": 0.09284202754497528, + "qparams": { + "group_size": 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": 0.08908134698867798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08045675605535507, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.046603601425886154, + "qparams": { + "group_size": 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": 0.04342658445239067, + "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": 0.05497446283698082, + "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": 0.05076802521944046, + "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": 0.047168415039777756, + "qparams": { + "group_size": 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": 0.041355859488248825, + "qparams": { + "group_size": 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": 0.039706259965896606, + "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": 0.02780618518590927, + "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": 0.02428598515689373, + "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": 0.022342391312122345, + "qparams": { + "group_size": 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": 0.021884793415665627, + "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": 0.013886310160160065, + "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": 0.011420144699513912, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011077377013862133, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.010350079275667667, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010038129054009914, + "qparams": { + "group_size": 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": 0.0072361077181994915, + "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": 0.0070303441025316715, + "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": 0.006638487335294485, + "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": 0.0043439180590212345, + "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": 0.09939263015985489, + "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": 0.09939263015985489, + "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": 0.2469831109046936, + "qparams": { + "group_size": 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": 0.2305615246295929, + "qparams": { + "group_size": 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": 0.22414976358413696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.20250935852527618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.11635148525238037, + "qparams": { + "group_size": 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": 0.10996399074792862, + "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": 0.13196951150894165, + "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": 0.12136925011873245, + "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": 0.1177155151963234, + "qparams": { + "group_size": 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": 0.10283910483121872, + "qparams": { + "group_size": 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": 0.09841890633106232, + "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": 0.06722392141819, + "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": 0.05805666744709015, + "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": 0.055629994720220566, + "qparams": { + "group_size": 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": 0.05507444962859154, + "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": 0.03363784775137901, + "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": 0.028147613629698753, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.027624843642115593, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02532299794256687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024937009438872337, + "qparams": { + "group_size": 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": 0.01720310188829899, + "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": 0.016453329473733902, + "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": 0.01624109037220478, + "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": 0.009765373542904854, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option_max": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09841890633106232, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + "best_option": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09841890633106232, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "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": 0.18154969811439514, + "qparams": { + "group_size": 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": 0.15423689782619476, + "qparams": { + "group_size": 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": 0.1444651186466217, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12413108348846436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08403301984071732, + "qparams": { + "group_size": 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": 0.07365667074918747, + "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": 0.09822732955217361, + "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": 0.09012535214424133, + "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": 0.08626871556043625, + "qparams": { + "group_size": 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": 0.0653228759765625, + "qparams": { + "group_size": 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": 0.06176912039518356, + "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": 0.05052719637751579, + "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": 0.043955009430646896, + "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": 0.04121841862797737, + "qparams": { + "group_size": 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": 0.04054855927824974, + "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": 0.02558911219239235, + "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": 0.022806817665696144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022188259288668633, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01947772316634655, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019050728529691696, + "qparams": { + "group_size": 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": 0.014511494897305965, + "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": 0.01616104505956173, + "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": 0.013715564273297787, + "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": 0.012906993739306927, + "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": 0.09822732955217361, + "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": 0.09822732955217361, + "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": 0.19014659523963928, + "qparams": { + "group_size": 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": 0.17993897199630737, + "qparams": { + "group_size": 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": 0.1765916496515274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16151204705238342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08984766900539398, + "qparams": { + "group_size": 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": 0.08621831238269806, + "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": 0.09945842623710632, + "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": 0.09208440035581589, + "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": 0.09055456519126892, + "qparams": { + "group_size": 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": 0.0811205580830574, + "qparams": { + "group_size": 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": 0.0778856948018074, + "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": 0.05058552324771881, + "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": 0.04400151968002319, + "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": 0.04298131540417671, + "qparams": { + "group_size": 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": 0.04273875430226326, + "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": 0.025253508239984512, + "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": 0.02197740226984024, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021725215017795563, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02017931081354618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020022835582494736, + "qparams": { + "group_size": 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": 0.013162786141037941, + "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": 0.013037797063589096, + "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": 0.01282407995313406, + "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": 0.00837732758373022, + "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": 0.09945842623710632, + "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": 0.09945842623710632, + "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": 0.2120392769575119, + "qparams": { + "group_size": 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": 0.20082662999629974, + "qparams": { + "group_size": 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": 0.19712267816066742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18028971552848816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10015419125556946, + "qparams": { + "group_size": 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": 0.09620112925767899, + "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": 0.11070913821458817, + "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": 0.10250966250896454, + "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": 0.1009702980518341, + "qparams": { + "group_size": 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": 0.09045641869306564, + "qparams": { + "group_size": 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": 0.08683544397354126, + "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": 0.056230127811431885, + "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": 0.04892793670296669, + "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": 0.04785726219415665, + "qparams": { + "group_size": 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": 0.04760601744055748, + "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": 0.028055988252162933, + "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": 0.024346210062503815, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024066073819994926, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02233235165476799, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.0221671462059021, + "qparams": { + "group_size": 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": 0.014492182061076164, + "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": 0.014235022477805614, + "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": 0.014132530428469181, + "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": 0.008877109736204147, + "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": 0.10015419125556946, + "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": 0.10015419125556946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.21385394036769867, + "qparams": { + "group_size": 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": 0.18081623315811157, + "qparams": { + "group_size": 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": 0.16641022264957428, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14355584979057312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0963125005364418, + "qparams": { + "group_size": 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": 0.08342530578374863, + "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": 0.12084277719259262, + "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": 0.10981705039739609, + "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": 0.10156900435686111, + "qparams": { + "group_size": 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": 0.07861915230751038, + "qparams": { + "group_size": 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": 0.0729815810918808, + "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": 0.061201248317956924, + "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": 0.05268530920147896, + "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": 0.04658886417746544, + "qparams": { + "group_size": 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": 0.045030444860458374, + "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": 0.030897121876478195, + "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": 0.024624213576316833, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02426806278526783, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.02120531164109707, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020061274990439415, + "qparams": { + "group_size": 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": 0.01690828613936901, + "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": 0.016594653949141502, + "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": 0.014965643174946308, + "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": 0.011574150063097477, + "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": 0.0963125005364418, + "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": 0.0963125005364418, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.1193951666355133, + "qparams": { + "group_size": 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": 0.11191818118095398, + "qparams": { + "group_size": 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": 0.108670175075531, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0985003113746643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05635816231369972, + "qparams": { + "group_size": 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": 0.053277499973773956, + "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": 0.06416890025138855, + "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": 0.05930466204881668, + "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": 0.05693698301911354, + "qparams": { + "group_size": 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": 0.05021803826093674, + "qparams": { + "group_size": 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": 0.04806334897875786, + "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": 0.03264758363366127, + "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": 0.028407247737050056, + "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": 0.027031298726797104, + "qparams": { + "group_size": 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": 0.026703614741563797, + "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": 0.01632980816066265, + "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": 0.0139002138748765, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.013635402545332909, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.012645049951970577, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01243508979678154, + "qparams": { + "group_size": 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": 0.008517169393599033, + "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": 0.00849786028265953, + "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": 0.008061079308390617, + "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": 0.005548133049160242, + "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": 0.0985003113746643, + "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": 0.0985003113746643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.k_proj", + "numel": 2097152, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1904296875, + "total_bits": 4593664.0, + "err": 0.10026903450489044, + "qparams": { + "group_size": 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": 0.09367253631353378, + "qparams": { + "group_size": 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": 0.08995676785707474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08125109225511551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.047148264944553375, + "qparams": { + "group_size": 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": 0.04402633383870125, + "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": 0.055188797414302826, + "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": 0.05100530758500099, + "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": 0.04767989367246628, + "qparams": { + "group_size": 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": 0.04186105355620384, + "qparams": { + "group_size": 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": 0.0400204174220562, + "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": 0.027989812195301056, + "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": 0.02444625459611416, + "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": 0.022611750289797783, + "qparams": { + "group_size": 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": 0.022158868610858917, + "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": 0.014011934399604797, + "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": 0.011547807604074478, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011212436482310295, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.010439885780215263, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010161546990275383, + "qparams": { + "group_size": 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": 0.007277360651642084, + "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": 0.007079016417264938, + "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": 0.006700514350086451, + "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": 0.0044034929014742374, + "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": 0.10026903450489044, + "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": 0.10026903450489044, + "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": 0.24931196868419647, + "qparams": { + "group_size": 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": 0.23213541507720947, + "qparams": { + "group_size": 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": 0.2259090095758438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.2040187120437622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.11739346385002136, + "qparams": { + "group_size": 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": 0.11085525155067444, + "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": 0.13284333050251007, + "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": 0.12209810316562653, + "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": 0.11866392195224762, + "qparams": { + "group_size": 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": 0.10353752225637436, + "qparams": { + "group_size": 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": 0.09882964938879013, + "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": 0.0675869733095169, + "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": 0.058344516903162, + "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": 0.05607110634446144, + "qparams": { + "group_size": 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": 0.05552908778190613, + "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": 0.033841006457805634, + "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": 0.02833128720521927, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.027839800342917442, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.025460587814450264, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02508830651640892, + "qparams": { + "group_size": 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": 0.017267484217882156, + "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": 0.01651209406554699, + "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": 0.01632712222635746, + "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": 0.009739862754940987, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option_max": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09882964938879013, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + "best_option": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09882964938879013, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "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": 0.12088460475206375, + "qparams": { + "group_size": 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": 0.09557390958070755, + "qparams": { + "group_size": 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": 0.0847942903637886, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07869288325309753, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.055033888667821884, + "qparams": { + "group_size": 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": 0.04384807497262955, + "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": 0.06923221051692963, + "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": 0.06324336677789688, + "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": 0.057884376496076584, + "qparams": { + "group_size": 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": 0.04291917383670807, + "qparams": { + "group_size": 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": 0.04166022315621376, + "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": 0.03576090931892395, + "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": 0.031284548342227936, + "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": 0.027521168813109398, + "qparams": { + "group_size": 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": 0.026564551517367363, + "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": 0.018280671909451485, + "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": 0.015907052904367447, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.015068494714796543, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.013883394189178944, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.013310606591403484, + "qparams": { + "group_size": 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": 0.010484599508345127, + "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": 0.012187845073640347, + "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": 0.009382941760122776, + "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": 0.010079668834805489, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09557390958070755, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09557390958070755, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 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": 0.18766288459300995, + "qparams": { + "group_size": 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": 0.17756056785583496, + "qparams": { + "group_size": 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": 0.17415082454681396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.159040629863739, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0888451412320137, + "qparams": { + "group_size": 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": 0.08515042066574097, + "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": 0.09855343401432037, + "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": 0.09113548696041107, + "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": 0.08954811096191406, + "qparams": { + "group_size": 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": 0.08007298409938812, + "qparams": { + "group_size": 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": 0.07684994488954544, + "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": 0.05016128718852997, + "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": 0.04356745630502701, + "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": 0.04251726344227791, + "qparams": { + "group_size": 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": 0.04226299375295639, + "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": 0.025080936029553413, + "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": 0.021772587671875954, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.021518170833587646, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01998330093920231, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019822215661406517, + "qparams": { + "group_size": 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": 0.01317153312265873, + "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": 0.01297720242291689, + "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": 0.012826588936150074, + "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": 0.008404488675296307, + "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": 0.09855343401432037, + "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": 0.09855343401432037, + "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": 0.20828339457511902, + "qparams": { + "group_size": 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": 0.1970423310995102, + "qparams": { + "group_size": 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": 0.1933813989162445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.17659732699394226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09843520075082779, + "qparams": { + "group_size": 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": 0.09442281723022461, + "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": 0.1090199276804924, + "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": 0.1008298397064209, + "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": 0.0992124080657959, + "qparams": { + "group_size": 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": 0.08875363320112228, + "qparams": { + "group_size": 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": 0.0850750282406807, + "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": 0.0553930327296257, + "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": 0.048151250928640366, + "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": 0.04705674946308136, + "qparams": { + "group_size": 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": 0.046792544424533844, + "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": 0.02766561135649681, + "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": 0.023945648223161697, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023669075220823288, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021941782906651497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02177364192903042, + "qparams": { + "group_size": 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": 0.014364928938448429, + "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": 0.014032069593667984, + "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": 0.013987021520733833, + "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": 0.008771862834692001, + "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": 3.1876362645348837, + "total_bits": 143726592.0, + "err": 0.0992124080657959, + "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": 3.2188862645348837, + "total_bits": 145135616.0, + "err": 0.09843520075082779, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.2175641655921936, + "qparams": { + "group_size": 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": 0.1849663406610489, + "qparams": { + "group_size": 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": 0.17091679573059082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.14751297235488892, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09817203879356384, + "qparams": { + "group_size": 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": 0.08561064302921295, + "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": 0.12212472409009933, + "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": 0.11128375679254532, + "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": 0.10331631451845169, + "qparams": { + "group_size": 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": 0.08043304830789566, + "qparams": { + "group_size": 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": 0.0746413916349411, + "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": 0.06215692684054375, + "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": 0.05344443395733833, + "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": 0.04747072234749794, + "qparams": { + "group_size": 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": 0.045956116169691086, + "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": 0.03152124211192131, + "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": 0.025076033547520638, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.024722589179873466, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021637242287397385, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020513642579317093, + "qparams": { + "group_size": 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": 0.01743914559483528, + "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": 0.01683337800204754, + "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": 0.015563355758786201, + "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": 0.01173569355159998, + "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": 0.09817203879356384, + "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": 0.09817203879356384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "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": 0.11230842769145966, + "qparams": { + "group_size": 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": 0.10520754754543304, + "qparams": { + "group_size": 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": 0.10206994414329529, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09248507022857666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05307397618889809, + "qparams": { + "group_size": 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": 0.0501415990293026, + "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": 0.06072401627898216, + "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": 0.05598035454750061, + "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": 0.053626276552677155, + "qparams": { + "group_size": 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": 0.04724802449345589, + "qparams": { + "group_size": 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": 0.045285001397132874, + "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": 0.03090214543044567, + "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": 0.026798736304044724, + "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": 0.025472646579146385, + "qparams": { + "group_size": 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": 0.025150757282972336, + "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": 0.015461832284927368, + "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": 0.013073309324681759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012811822816729546, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011878845281898975, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011673558503389359, + "qparams": { + "group_size": 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": 0.008074501529335976, + "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": 0.007965004071593285, + "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": 0.007628465071320534, + "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": 0.005143079441040754, + "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": 0.09248507022857666, + "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": 0.09248507022857666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.k_proj", + "numel": 2097152, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1904296875, + "total_bits": 4593664.0, + "err": 0.09377618134021759, + "qparams": { + "group_size": 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": 0.08751610666513443, + "qparams": { + "group_size": 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": 0.08397839218378067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07592112571001053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04410781338810921, + "qparams": { + "group_size": 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": 0.041095856577157974, + "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": 0.05176888033747673, + "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": 0.047829560935497284, + "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": 0.04458210989832878, + "qparams": { + "group_size": 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": 0.03908565640449524, + "qparams": { + "group_size": 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": 0.037368856370449066, + "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": 0.026224792003631592, + "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": 0.02291952073574066, + "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": 0.02118200622498989, + "qparams": { + "group_size": 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": 0.020739197731018066, + "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": 0.013101406395435333, + "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": 0.010815849527716637, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010511468164622784, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.009791120886802673, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.009521985426545143, + "qparams": { + "group_size": 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": 0.006818066816776991, + "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": 0.00665039150044322, + "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": 0.006273283623158932, + "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": 0.004165417980402708, + "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": 0.09377618134021759, + "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": 0.09377618134021759, + "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": 0.24394510686397552, + "qparams": { + "group_size": 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": 0.22727376222610474, + "qparams": { + "group_size": 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": 0.2209790050983429, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.19934023916721344, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.11487352102994919, + "qparams": { + "group_size": 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": 0.10862577706575394, + "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": 0.13011910021305084, + "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": 0.11986886709928513, + "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": 0.1162281483411789, + "qparams": { + "group_size": 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": 0.10130796581506729, + "qparams": { + "group_size": 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": 0.09667326509952545, + "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": 0.06624811142683029, + "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": 0.057273682206869125, + "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": 0.05495745316147804, + "qparams": { + "group_size": 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": 0.054441072046756744, + "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": 0.03311409056186676, + "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": 0.027863003313541412, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02735661342740059, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.025045305490493774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02467159368097782, + "qparams": { + "group_size": 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": 0.01693517155945301, + "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": 0.01636425033211708, + "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": 0.016031909734010696, + "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": 0.009834742173552513, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option_max": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09667326509952545, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + "best_option": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.65771484375, + "total_bits": 7670784.0, + "err": 0.09667326509952545, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "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": 0.11920561641454697, + "qparams": { + "group_size": 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": 0.10379314422607422, + "qparams": { + "group_size": 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": 0.09875218570232391, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08399659395217896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.05584144592285156, + "qparams": { + "group_size": 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": 0.05066671594977379, + "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": 0.06397244334220886, + "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": 0.0584915429353714, + "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": 0.05684184283018112, + "qparams": { + "group_size": 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": 0.04422670975327492, + "qparams": { + "group_size": 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": 0.04126817733049393, + "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": 0.032955918461084366, + "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": 0.028506364673376083, + "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": 0.027307609096169472, + "qparams": { + "group_size": 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": 0.027020111680030823, + "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": 0.016663087531924248, + "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": 0.014954368583858013, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.014656312763690948, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.012840762734413147, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01265254057943821, + "qparams": { + "group_size": 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": 0.009377140551805496, + "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": 0.010333934798836708, + "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": 0.009016064926981926, + "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": 0.008191102184355259, + "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": 0.09875218570232391, + "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": 0.09875218570232391, + "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": 0.17836306989192963, + "qparams": { + "group_size": 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": 0.1686599850654602, + "qparams": { + "group_size": 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": 0.16527727246284485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.15093635022640228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0846959576010704, + "qparams": { + "group_size": 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": 0.08116044104099274, + "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": 0.09411732107400894, + "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": 0.08700627833604813, + "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": 0.08540999889373779, + "qparams": { + "group_size": 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": 0.07627946138381958, + "qparams": { + "group_size": 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": 0.0731927901506424, + "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": 0.048076435923576355, + "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": 0.04176328703761101, + "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": 0.04072475805878639, + "qparams": { + "group_size": 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": 0.04047317057847977, + "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": 0.02405102364718914, + "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": 0.021194348111748695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.020950552076101303, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01951085962355137, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019357766956090927, + "qparams": { + "group_size": 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": 0.012842093594372272, + "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": 0.013149762526154518, + "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": 0.012509776279330254, + "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": 0.009182380512356758, + "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": 0.09411732107400894, + "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": 0.09411732107400894, + "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": 0.18908259272575378, + "qparams": { + "group_size": 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": 0.17870599031448364, + "qparams": { + "group_size": 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": 0.17529122531414032, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.16007006168365479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09003656357526779, + "qparams": { + "group_size": 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": 0.0863495022058487, + "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": 0.09985106438398361, + "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": 0.09231237322092056, + "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": 0.0907742977142334, + "qparams": { + "group_size": 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": 0.08117993921041489, + "qparams": { + "group_size": 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": 0.07800135761499405, + "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": 0.05143074691295624, + "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": 0.04468366503715515, + "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": 0.043671827763319016, + "qparams": { + "group_size": 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": 0.0434313528239727, + "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": 0.02580348588526249, + "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": 0.023364203050732613, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023117655888199806, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.021664660423994064, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02151147648692131, + "qparams": { + "group_size": 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": 0.014252692461013794, + "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": 0.015345459803938866, + "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": 0.013923367485404015, + "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": 0.011664923280477524, + "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": 0.09985106438398361, + "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": 0.09985106438398361, + "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": 0.2029201239347458, + "qparams": { + "group_size": 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": 0.1727362424135208, + "qparams": { + "group_size": 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": 0.15943574905395508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.1376928836107254, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09206458926200867, + "qparams": { + "group_size": 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": 0.08006427437067032, + "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": 0.11426414549350739, + "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": 0.10454603284597397, + "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": 0.096722811460495, + "qparams": { + "group_size": 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": 0.0754585936665535, + "qparams": { + "group_size": 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": 0.07003170251846313, + "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": 0.058532264083623886, + "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": 0.05054106563329697, + "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": 0.04479306563735008, + "qparams": { + "group_size": 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": 0.043321747332811356, + "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": 0.029708093032240868, + "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": 0.02411048859357834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023761838674545288, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.020981526002287865, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019947631284594536, + "qparams": { + "group_size": 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": 0.016477925702929497, + "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": 0.016729533672332764, + "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": 0.014676828868687153, + "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": 0.012308378703892231, + "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": 0.096722811460495, + "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": 0.096722811460495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "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": 0.10827305167913437, + "qparams": { + "group_size": 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": 0.1020360141992569, + "qparams": { + "group_size": 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": 0.09936682879924774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09043831378221512, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.051272377371788025, + "qparams": { + "group_size": 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": 0.0487651601433754, + "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": 0.05817551910877228, + "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": 0.05361885949969292, + "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": 0.05172334611415863, + "qparams": { + "group_size": 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": 0.04596276208758354, + "qparams": { + "group_size": 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": 0.04414921626448631, + "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": 0.02961832471191883, + "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": 0.025650661438703537, + "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": 0.024618318304419518, + "qparams": { + "group_size": 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": 0.02436506748199463, + "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": 0.01484359335154295, + "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": 0.012678369879722595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.012465490028262138, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011598394252359867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011440727859735489, + "qparams": { + "group_size": 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": 0.007819794118404388, + "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": 0.007735037244856358, + "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": 0.007477365434169769, + "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": 0.005107318516820669, + "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": 0.09936682879924774, + "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": 0.09936682879924774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "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": 0.0985112115740776, + "qparams": { + "group_size": 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": 0.09271882474422455, + "qparams": { + "group_size": 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": 0.08990991860628128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0815734788775444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.046510256826877594, + "qparams": { + "group_size": 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": 0.043963607400655746, + "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": 0.05320625379681587, + "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": 0.049203477799892426, + "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": 0.046897608786821365, + "qparams": { + "group_size": 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": 0.04156138747930527, + "qparams": { + "group_size": 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": 0.039843086153268814, + "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": 0.02696368284523487, + "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": 0.023498551920056343, + "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": 0.022282525897026062, + "qparams": { + "group_size": 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": 0.021965643391013145, + "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": 0.013490867801010609, + "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": 0.0113284382969141, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.011086860671639442, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.01030284259468317, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010114265605807304, + "qparams": { + "group_size": 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": 0.006970375310629606, + "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": 0.006779191084206104, + "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": 0.006566469557583332, + "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": 0.004160519689321518, + "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": 0.0985112115740776, + "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": 0.0985112115740776, + "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": 0.23123064637184143, + "qparams": { + "group_size": 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": 0.21768994629383087, + "qparams": { + "group_size": 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": 0.21340098977088928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.19368204474449158, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10913192480802536, + "qparams": { + "group_size": 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": 0.1042238175868988, + "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": 0.121273934841156, + "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": 0.11197388172149658, + "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": 0.11010304093360901, + "qparams": { + "group_size": 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": 0.097552590072155, + "qparams": { + "group_size": 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": 0.0931277945637703, + "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": 0.06174282729625702, + "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": 0.053506042808294296, + "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": 0.05208274722099304, + "qparams": { + "group_size": 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": 0.05177072435617447, + "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": 0.03085670992732048, + "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": 0.026319582015275955, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02597329579293728, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.023895179852843285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.023674821481108665, + "qparams": { + "group_size": 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": 0.015749309211969376, + "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": 0.015126790851354599, + "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": 0.015171974897384644, + "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": 0.008903619833290577, + "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": 0.097552590072155, + "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": 0.097552590072155, + "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": 0.16461214423179626, + "qparams": { + "group_size": 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": 0.14576491713523865, + "qparams": { + "group_size": 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": 0.13953521847724915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.11831098794937134, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07653093338012695, + "qparams": { + "group_size": 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": 0.07006560266017914, + "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": 0.08747542649507523, + "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": 0.08047477155923843, + "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": 0.07789156585931778, + "qparams": { + "group_size": 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": 0.061932217329740524, + "qparams": { + "group_size": 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": 0.05704836547374725, + "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": 0.045042864978313446, + "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": 0.03891758620738983, + "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": 0.03716559708118439, + "qparams": { + "group_size": 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": 0.036740295588970184, + "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": 0.022647056728601456, + "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": 0.01994483545422554, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.019537512212991714, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.017197128385305405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.016923440620303154, + "qparams": { + "group_size": 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": 0.012503732927143574, + "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": 0.013320346362888813, + "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": 0.011963659897446632, + "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": 0.01010812446475029, + "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": 0.08747542649507523, + "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": 0.08747542649507523, + "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": 0.16398370265960693, + "qparams": { + "group_size": 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": 0.1549561470746994, + "qparams": { + "group_size": 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": 0.15178410708904266, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.13857005536556244, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07786686718463898, + "qparams": { + "group_size": 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": 0.07452841848134995, + "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": 0.08660749346017838, + "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": 0.08009400218725204, + "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": 0.07851645350456238, + "qparams": { + "group_size": 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": 0.07009825855493546, + "qparams": { + "group_size": 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": 0.06725353747606277, + "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": 0.04431869462132454, + "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": 0.03846283629536629, + "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": 0.037436313927173615, + "qparams": { + "group_size": 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": 0.037189483642578125, + "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": 0.022157734259963036, + "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": 0.019493786618113518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01925126276910305, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.017937343567609787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017778055742383003, + "qparams": { + "group_size": 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": 0.011805115267634392, + "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": 0.01210707426071167, + "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": 0.011466076597571373, + "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": 0.008447709493339062, + "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": 0.08660749346017838, + "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": 0.08660749346017838, + "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": 0.16044625639915466, + "qparams": { + "group_size": 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": 0.15165995061397552, + "qparams": { + "group_size": 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": 0.14859168231487274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.13563452661037445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.07607439160346985, + "qparams": { + "group_size": 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": 0.07288264483213425, + "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": 0.08458062261343002, + "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": 0.0781649574637413, + "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": 0.07670394331216812, + "qparams": { + "group_size": 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": 0.06846409291028976, + "qparams": { + "group_size": 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": 0.06566178798675537, + "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": 0.043075453490018845, + "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": 0.03740837797522545, + "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": 0.036457423120737076, + "qparams": { + "group_size": 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": 0.036232367157936096, + "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": 0.02151772752404213, + "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": 0.01875929720699787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01853141561150551, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.017209211364388466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01707136258482933, + "qparams": { + "group_size": 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": 0.011276676319539547, + "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": 0.011321867816150188, + "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": 0.010947801172733307, + "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": 0.007505912333726883, + "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": 0.08458062261343002, + "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": 0.08458062261343002, + "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": 0.14178533852100372, + "qparams": { + "group_size": 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": 0.12082989513874054, + "qparams": { + "group_size": 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": 0.11102333664894104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.09608351439237595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.0644257441163063, + "qparams": { + "group_size": 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": 0.05589067190885544, + "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": 0.08133592456579208, + "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": 0.0740736722946167, + "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": 0.0677293911576271, + "qparams": { + "group_size": 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": 0.05303123965859413, + "qparams": { + "group_size": 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": 0.04954637587070465, + "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": 0.04159345105290413, + "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": 0.035899076610803604, + "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": 0.031553301960229874, + "qparams": { + "group_size": 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": 0.030454091727733612, + "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": 0.020994506776332855, + "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": 0.017322344705462456, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017032677307724953, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.015233923681080341, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.014484316110610962, + "qparams": { + "group_size": 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": 0.011625031009316444, + "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": 0.012416146695613861, + "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": 0.01025288738310337, + "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": 0.009507762268185616, + "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": 2.730017373728198, + "total_bits": 123093120.00000001, + "err": 0.09608351439237595, + "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": 2.730017373728198, + "total_bits": 123093120.00000001, + "err": 0.09608351439237595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.q_proj", + "numel": 16777216, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1878662109375, + "total_bits": 36706304.0, + "err": 0.10350323468446732, + "qparams": { + "group_size": 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": 0.09748492389917374, + "qparams": { + "group_size": 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": 0.09501852840185165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08648636937141418, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04911094531416893, + "qparams": { + "group_size": 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": 0.04669025540351868, + "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": 0.055848874151706696, + "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": 0.05121003836393356, + "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": 0.04954960569739342, + "qparams": { + "group_size": 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": 0.04400403052568436, + "qparams": { + "group_size": 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": 0.04233361780643463, + "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": 0.028443468734622, + "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": 0.02450052835047245, + "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": 0.023556021973490715, + "qparams": { + "group_size": 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": 0.02333204448223114, + "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": 0.0142352394759655, + "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": 0.012067299336194992, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01186703983694315, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.011012810282409191, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010867411270737648, + "qparams": { + "group_size": 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": 0.007414595223963261, + "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": 0.007252716924995184, + "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": 0.0070955390110611916, + "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": 0.004654204472899437, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09748492389917374, + "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": 2.3753662109375, + "total_bits": 39852032.0, + "err": 0.09748492389917374, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "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": 0.09686928987503052, + "qparams": { + "group_size": 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": 0.09110111743211746, + "qparams": { + "group_size": 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": 0.08849728107452393, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.08022495359182358, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04568175971508026, + "qparams": { + "group_size": 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": 0.043234653770923615, + "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": 0.052050746977329254, + "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": 0.04820253700017929, + "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": 0.04609799012541771, + "qparams": { + "group_size": 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": 0.040822919458150864, + "qparams": { + "group_size": 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": 0.03901147469878197, + "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": 0.026378067210316658, + "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": 0.023046202957630157, + "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": 0.021899143233895302, + "qparams": { + "group_size": 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": 0.0216146819293499, + "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": 0.01318155787885189, + "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": 0.011122959665954113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010900317691266537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.010091934353113174, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00993000715970993, + "qparams": { + "group_size": 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": 0.006818968337029219, + "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": 0.006608455441892147, + "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": 0.006451358553022146, + "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": 0.0040360852144658566, + "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": 0.09686928987503052, + "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": 0.09686928987503052, + "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": 0.21684344112873077, + "qparams": { + "group_size": 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": 0.20375895500183105, + "qparams": { + "group_size": 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": 0.19936691224575043, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.18081237375736237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10234663635492325, + "qparams": { + "group_size": 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": 0.09761360287666321, + "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": 0.11451900750398636, + "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": 0.10546733438968658, + "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": 0.10338144749403, + "qparams": { + "group_size": 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": 0.09135313332080841, + "qparams": { + "group_size": 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": 0.08736690878868103, + "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": 0.05829303339123726, + "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": 0.050359342247247696, + "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": 0.048915233463048935, + "qparams": { + "group_size": 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": 0.048594195395708084, + "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": 0.029057873412966728, + "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": 0.024693885818123817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.02434803731739521, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.022397726774215698, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.022150028496980667, + "qparams": { + "group_size": 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": 0.014850341714918613, + "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": 0.014220716431736946, + "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": 0.014274891465902328, + "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": 0.008388442918658257, + "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": 0.09135313332080841, + "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": 0.09135313332080841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "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": 0.058811403810977936, + "qparams": { + "group_size": 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": 0.0512092150747776, + "qparams": { + "group_size": 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": 0.04841671511530876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.04022867977619171, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.02733241580426693, + "qparams": { + "group_size": 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": 0.024357350543141365, + "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": 0.031767409294843674, + "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": 0.02906375378370285, + "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": 0.02791200578212738, + "qparams": { + "group_size": 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": 0.021828966215252876, + "qparams": { + "group_size": 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": 0.019622409716248512, + "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": 0.016466638073325157, + "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": 0.01460955012589693, + "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": 0.013841398991644382, + "qparams": { + "group_size": 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": 0.013663575984537601, + "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": 0.00871703214943409, + "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": 0.008260713890194893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.00808311440050602, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.007370860315859318, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.007268417160958052, + "qparams": { + "group_size": 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": 0.005461479537189007, + "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": 0.006401547696441412, + "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": 0.005266326479613781, + "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": 0.005587765481323004, + "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": 0.058811403810977936, + "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": 0.058811403810977936, + "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": 0.12972956895828247, + "qparams": { + "group_size": 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": 0.1226426213979721, + "qparams": { + "group_size": 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": 0.11954478174448013, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.10903232544660568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.062222886830568314, + "qparams": { + "group_size": 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": 0.059241607785224915, + "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": 0.07118216902017593, + "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": 0.06516951322555542, + "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": 0.06271912902593613, + "qparams": { + "group_size": 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": 0.0561027005314827, + "qparams": { + "group_size": 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": 0.05408351123332977, + "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": 0.036758072674274445, + "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": 0.032146114856004715, + "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": 0.0308234803378582, + "qparams": { + "group_size": 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": 0.03050430305302143, + "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": 0.018702520057559013, + "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": 0.017398806288838387, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.017156383022665977, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.016306346282362938, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01612946018576622, + "qparams": { + "group_size": 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": 0.010821432806551456, + "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": 0.01255510188639164, + "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": 0.010398291051387787, + "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": 0.010416180826723576, + "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": 0.07118216902017593, + "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": 0.07118216902017593, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "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": 0.14482849836349487, + "qparams": { + "group_size": 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": 0.13698837161064148, + "qparams": { + "group_size": 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": 0.13338086009025574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.12175268679857254, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.06931943446397781, + "qparams": { + "group_size": 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": 0.06611189246177673, + "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": 0.0790897011756897, + "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": 0.07251143455505371, + "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": 0.06988707184791565, + "qparams": { + "group_size": 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": 0.06257597357034683, + "qparams": { + "group_size": 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": 0.0603192001581192, + "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": 0.04085652902722359, + "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": 0.03584848344326019, + "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": 0.03452017530798912, + "qparams": { + "group_size": 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": 0.03418688103556633, + "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": 0.020894769579172134, + "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": 0.019755415618419647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01949971728026867, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.018580326810479164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.01838880218565464, + "qparams": { + "group_size": 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": 0.01232067123055458, + "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": 0.014526668936014175, + "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": 0.011840789578855038, + "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": 0.012242852710187435, + "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": 0.0790897011756897, + "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": 0.0790897011756897, + "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": 0.06918499618768692, + "qparams": { + "group_size": 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": 0.059555478394031525, + "qparams": { + "group_size": 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": 0.04977088049054146, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.043774429708719254, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 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": 0.03178466111421585, + "qparams": { + "group_size": 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": 0.025778764858841896, + "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": 0.04777919501066208, + "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": 0.04329176992177963, + "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": 0.03344021365046501, + "qparams": { + "group_size": 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": 0.027034714818000793, + "qparams": { + "group_size": 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": 0.026623554527759552, + "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": 0.02492288127541542, + "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": 0.02123214304447174, + "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": 0.01712064817547798, + "qparams": { + "group_size": 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": 0.0158973578363657, + "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": 0.013397554866969585, + "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": 0.011359630152583122, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010995193384587765, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 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": 0.010709427297115326, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 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": 0.010177656076848507, + "qparams": { + "group_size": 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": 0.008781291544437408, + "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": 0.010004906915128231, + "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": 0.007663393393158913, + "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": 0.008981546387076378, + "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": 0.06918499618768692, + "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": 0.06918499618768692, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + } + ], + "last_module_idx": 66, + "base_perplexity": 9.43343121924492 +} \ No newline at end of file