{ "measurement": [ { "key": "model.layers.0.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.02171376906335354, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.02152251824736595, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.011157726868987083, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.011104918085038662, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.011047903448343277, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.007671317085623741, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.036719925701618195, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.021320758387446404, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.011063977144658566, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.011024624109268188, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.01117617916315794, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.013199222274124622, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.011010157875716686, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.008365418761968613, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.00764921959489584, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.008415093645453453, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.007643500342965126, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.007381145376712084, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.0076424372382462025, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.00738042825832963, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0075923241674900055, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007642253767699003, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007297586649656296, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0073792897164821625, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.02171376906335354, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.02171376906335354, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.022244811058044434, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.021980799734592438, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.010306116193532944, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.010237591341137886, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.010166866704821587, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.005353022366762161, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.02991025149822235, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.02174502983689308, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.010197301395237446, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.010127966292202473, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.010235493071377277, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.011110011488199234, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.010107462294399738, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.00651503074914217, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.005313773639500141, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.006346100475639105, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.005299089942127466, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.004882996901869774, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.005296424031257629, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.004879907239228487, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.004896655213087797, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.005295617505908012, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.004372746217995882, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00487753190100193, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.022244811058044434, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.022244811058044434, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.07380950450897217, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.06434576213359833, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.05652786046266556, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.04701913148164749, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.033771663904190063, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0276859812438488, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.054002709686756134, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.042861662805080414, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.035025276243686676, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.02772747538983822, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.027127213776111603, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02883334644138813, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.020838424563407898, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.01654069684445858, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.015342782251536846, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014685170724987984, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009007064625620842, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.008416480384767056, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.00788616668432951, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0071089742705225945, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007511761039495468, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0067228032276034355, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005092450883239508, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004779632203280926, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07380950450897217, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07380950450897217, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.0.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.02310180477797985, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.014769116416573524, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.009337054565548897, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.010239223949611187, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.009510375559329987, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.005094630643725395, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.015221863985061646, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.013373562134802341, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.01065692026168108, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.006990829017013311, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.007277282420545816, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.007663476746529341, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.006692772265523672, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.005100560840219259, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.004644818603992462, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.004061277490109205, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0034714876674115658, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.003337213769555092, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.003164989873766899, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0029575133230537176, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.002592545235529542, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.003134567989036441, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0021783653646707535, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0028091983404010534, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.02310180477797985, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.02310180477797985, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.056745532900094986, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.05366590991616249, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.05264659598469734, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.04834527149796486, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.025261133909225464, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.02434694953262806, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.028078468516469002, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.025963859632611275, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.025547970086336136, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.023362191393971443, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.022527989000082016, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.01422698050737381, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.012497185729444027, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.01220108475536108, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.012131311930716038, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.007176232989877462, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.006694160867482424, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.006671661511063576, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.00635997299104929, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.006313247140496969, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.00405222550034523, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.004603146109730005, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.003950257785618305, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.003681362606585026, "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": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.056745532900094986, "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": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.056745532900094986, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.08386430889368057, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.0794035792350769, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.07797013968229294, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.07169152796268463, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.03731698542833328, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.035993728786706924, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.041389767080545425, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.03824813663959503, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.037725869566202164, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.03450826555490494, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.033260278403759, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.020845746621489525, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.01812807098031044, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.017732102423906326, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.017635617405176163, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.010409540496766567, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.00923210196197033, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.009196102619171143, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.008694497868418694, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.00863788090646267, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.00555490143597126, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.005739763379096985, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.005411307327449322, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.004031873773783445, "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": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.08386430889368057, "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": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.08386430889368057, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.06312497705221176, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.0539185106754303, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.03390002250671387, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.031192556023597717, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.02608116716146469, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.01658027619123459, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.05284803360700607, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.04731897637248039, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.027746830135583878, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.023040423169732094, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.022884361445903778, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.02311556600034237, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.020994046702980995, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.014239145442843437, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.011938772164285183, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.012724602594971657, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.009539849124848843, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.008360030129551888, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.009140427224338055, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.0077857207506895065, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.008980032056570053, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.00888826884329319, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.007287578657269478, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.007244926877319813, "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": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.06312497705221176, "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": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.06312497705221176, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.013941861689090729, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.011173968203365803, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.008630888536572456, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.00757473474368453, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0059570311568677425, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0040461779572069645, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.010922110639512539, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.008646879345178604, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.006467272061854601, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.004946793429553509, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.005218397825956345, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.005744225811213255, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.004268473945558071, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.002985467901453376, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0025732857175171375, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.002962288446724415, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0017092579510062933, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.0015243568923324347, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.001541045494377613, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.001296297530643642, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0015947141218930483, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.001436307211406529, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0010038684122264385, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.001053700689226389, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.013941861689090729, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.013941861689090729, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.013073616661131382, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.010280197486281395, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.008001306094229221, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.00699690030887723, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.005447720643132925, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.003785650245845318, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.008896918967366219, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.007824558764696121, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0059775710105896, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.004420627374202013, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0044960686936974525, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.00450708856806159, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.003747753333300352, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.002691899659112096, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0023812749423086643, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.002274705795571208, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0015264456160366535, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.0013948521809652448, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.0013515056343749166, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0011619599536061287, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0012379818363115191, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0012457151897251606, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0008707621600478888, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0009170372504740953, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.013073616661131382, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.013073616661131382, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.07277064770460129, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.06192316114902496, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.057300228625535965, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.04946179687976837, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.032457102090120316, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.027948016300797462, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04016989469528198, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03684615716338158, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.034026723355054855, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.02655472606420517, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.02487124688923359, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.020516829565167427, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.017681801691651344, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.01565449871122837, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.015151039697229862, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.010297570377588272, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.008262727409601212, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.00807436928153038, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.007122586481273174, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0067896596156060696, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005433426238596439, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.005547836422920227, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0047221919521689415, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0038587504532188177, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07277064770460129, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07277064770460129, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.1.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1437220722436905, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0868394672870636, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.0576532781124115, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06311436742544174, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06199329346418381, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03532871976494789, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08675181865692139, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0768258199095726, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06834118068218231, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03912325203418732, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04137800261378288, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04460054263472557, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03793281689286232, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03127077594399452, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02953578531742096, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.022888757288455963, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.018545357510447502, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01800921931862831, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.015025487169623375, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01391570083796978, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013375268317759037, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014869365841150284, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.011406981386244297, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012336650863289833, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0868394672870636, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0868394672870636, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.10040490329265594, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.09582118690013885, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.09446932375431061, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.08718003332614899, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.044222358614206314, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.043015897274017334, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.04863402619957924, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.04497355967760086, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.04458824172616005, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.04134689271450043, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.04013906046748161, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.024610696360468864, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.02144099958240986, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.02112535573542118, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.021056829020380974, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.012343479320406914, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.011315414682030678, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.011294192634522915, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.010804837569594383, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.010761997662484646, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.006844777148216963, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.0074531808495521545, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.0067486329935491085, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.005731718149036169, "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": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.09582118690013885, "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": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.09582118690013885, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.12722691893577576, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.12179888784885406, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.12021947652101517, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.11137416958808899, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.05627055838704109, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.054790593683719635, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0617363415658474, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.057133082300424576, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.05669635534286499, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.052724700421094894, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.05110340565443039, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.030921878293156624, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.026892656460404396, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.026521388441324234, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.026434941217303276, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.0153807969763875, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.013482197187840939, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.013454482890665531, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.012793419882655144, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.012739527970552444, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.007966314442455769, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.007873674854636192, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.007839163765311241, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.004991075024008751, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0617363415658474, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0617363415658474, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.11181294918060303, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.09964095056056976, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.08586236089468002, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.07897496223449707, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.05072147026658058, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.03951745480298996, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.07311102002859116, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.06673412024974823, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.05245508998632431, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.046098742634058, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.044944219291210175, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.03544612228870392, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.03484059497714043, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.024382680654525757, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.022148258984088898, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.018994398415088654, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.014379456639289856, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.012957694940268993, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.013614458963274956, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.011959652416408062, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.012355620041489601, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.011852291412651539, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.009803158231079578, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.00885431282222271, "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": 2.62529296875, "total_bits": 185815296.0, "err": 0.08586236089468002, "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": 2.62529296875, "total_bits": 185815296.0, "err": 0.08586236089468002, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.2.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.030031180009245872, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.027650639414787292, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.026505112648010254, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.023759549483656883, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.013535398989915848, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.012530004605650902, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.017645511776208878, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.014728862792253494, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.013797352090477943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.012048564851284027, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.01151035912334919, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.008951584808528423, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.007071707863360643, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.006503957789391279, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.006362241227179766, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.004503375384956598, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.003431532997637987, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.003363851923495531, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.003145347349345684, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0030550817027688026, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.002404290484264493, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.002259758999571204, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.002051173010841012, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0015964406775310636, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.030031180009245872, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.030031180009245872, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.02438420243561268, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.022397221997380257, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.021482503041625023, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.019224682822823524, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.01094652246683836, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.010157078504562378, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.013030203990638256, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.011861500330269337, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.011175206862390041, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.009707321412861347, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.009222175925970078, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0065699610859155655, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.005646848119795322, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.005244446452707052, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.005146269220858812, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.003297538263723254, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0027426688466221094, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.0027006480377167463, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.0025010393001139164, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0024399026297032833, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0017515578074380755, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0017639382276684046, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0016035574954003096, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0012194691225886345, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.02438420243561268, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.02438420243561268, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11377263814210892, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10624870657920837, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10380075871944427, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09409904479980469, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05209032818675041, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04951637238264084, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05834628641605377, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05390855669975281, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05289428308606148, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04690925031900406, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04458560794591904, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.029535720124840736, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025634778663516045, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02481563575565815, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0246209055185318, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01473002228885889, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012642709538340569, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012572248466312885, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011583617888391018, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011459561996161938, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0075963446870446205, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0074750082567334175, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0072886454872787, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00469382805749774, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09409904479980469, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09409904479980469, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.2.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.14922590553760529, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12237802147865295, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11204883456230164, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08785771578550339, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06716469675302505, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05697578936815262, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08312267810106277, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0753808319568634, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07063906639814377, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.050709664821624756, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.045696891844272614, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04296470060944557, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.036996375769376755, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.033370375633239746, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03245305269956589, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.022074101492762566, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.019044190645217896, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01876402460038662, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01625078171491623, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.015703408047556877, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.012957131490111351, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014248420484364033, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.011897988617420197, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011723191477358341, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08785771578550339, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08785771578550339, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.14327971637248993, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1368204802274704, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.13497789204120636, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.12495571374893188, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0650559812784195, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.0631875991821289, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07159633934497833, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.06610042601823807, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.0656336173415184, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06060254946351051, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.05878305807709694, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.036298271268606186, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03147411346435547, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03102055937051773, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.030914776027202606, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.01815195381641388, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.016186852008104324, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.016160085797309875, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.0153351416811347, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.015275894664227962, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.009806211106479168, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.01007640827447176, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.009664509445428848, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007170502562075853, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07159633934497833, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07159633934497833, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1751893162727356, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16754132509231567, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.16534200310707092, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1532871127128601, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07978014647960663, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07750723510980606, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08763958513736725, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08104263246059418, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08046910166740417, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07437760382890701, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.07209974527359009, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04424170032143593, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03843221813440323, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03788015991449356, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.0377567857503891, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.022058334201574326, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01931120455265045, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01927429996430874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018228396773338318, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.018145762383937836, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.01151144877076149, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011343840509653091, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011336463503539562, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007265035063028336, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08763958513736725, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08763958513736725, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.16310428082942963, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.14574693143367767, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1388940066099167, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.12606672942638397, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.07288466393947601, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.06632459908723831, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.08631777763366699, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.07933669537305832, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.07551389932632446, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.06402231752872467, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.0615185871720314, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.04385285824537277, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.0379025824368, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.035007767379283905, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.03430194780230522, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02204030007123947, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.0184517540037632, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.018290940672159195, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.016752339899539948, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.01631687954068184, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.011992954649031162, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.012104679830372334, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.0110778184607625, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.00861331820487976, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.08631777763366699, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.08631777763366699, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.042518340051174164, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.03947439789772034, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.03797375038266182, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.034251537173986435, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.019407566636800766, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.018088631331920624, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0234677754342556, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.02094419300556183, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.01974162459373474, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.017384354025125504, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.016689317300915718, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.012027731165289879, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.010056018829345703, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.009325345046818256, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.009141206741333008, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.006088580936193466, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.004887774121016264, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.0048019192181527615, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.0044941105879843235, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0043810694478452206, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.003265653969720006, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.003167303279042244, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.002926459303125739, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0022077378816902637, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.042518340051174164, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.042518340051174164, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.034839365631341934, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.03232081979513168, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.03105621412396431, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.028000501915812492, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.015881292521953583, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.01482789870351553, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.01884111762046814, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.017151987180113792, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.01616009883582592, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.014212878420948982, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.013604752719402313, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.009523378685116768, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.008186334744095802, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.007630865089595318, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.007499544881284237, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.004776108544319868, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.004009697586297989, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.003949617501348257, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.003683974966406822, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0036005445290356874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.002553679049015045, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0025984784588217735, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0023639288265258074, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0018321851966902614, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.034839365631341934, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.034839365631341934, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.15479353070259094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14506804943084717, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.14191918075084686, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.12891720235347748, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07137925922870636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.06801410019397736, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08039674162864685, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07369109243154526, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07240848243236542, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06444115191698074, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.061495378613471985, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0407402478158474, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.035106685012578964, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.034056294709444046, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03379925712943077, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02034538984298706, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01735828071832657, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.017273563891649246, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.015941865742206573, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.015781940892338753, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.010560550726950169, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.010279933921992779, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.010153771378099918, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006490638945251703, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08039674162864685, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08039674162864685, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1338786631822586, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11274023354053497, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10299693048000336, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08181007206439972, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.060770608484745026, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05183122307062149, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0809100791811943, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0693502277135849, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06298614293336868, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04683011770248413, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.042707331478595734, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04127799719572067, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03469301387667656, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.030943121761083603, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.029947590082883835, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.022258583456277847, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01859409175813198, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01823483407497406, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.016350796446204185, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01581660658121109, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014205798506736755, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014706237241625786, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012942583300173283, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01265937089920044, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08181007206439972, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08181007206439972, "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": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1632927507162094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15530261397361755, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1529230773448944, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.14095330238342285, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07584503293037415, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07325813174247742, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08362041413784027, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07726974785327911, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07658091187477112, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06996992975473404, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06746622174978256, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04277260601520538, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.037286508828401566, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.036647651344537735, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.036493849009275436, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.021460536867380142, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.019660990685224533, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.019619861617684364, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018558520823717117, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01847623474895954, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011856446973979473, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.012974710203707218, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011654759757220745, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009974189102649689, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08362041413784027, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08362041413784027, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1999727338552475, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1903221607208252, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1874869167804718, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17275798320770264, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09260687977075577, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.08949664235115051, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.1021316647529602, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.0943874716758728, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09351743757724762, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0854392945766449, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0822838544845581, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05189474672079086, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.045009613037109375, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.044253915548324585, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04405532032251358, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.025926528498530388, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02286710776388645, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.022830093279480934, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02143876440823078, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.021346477791666985, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.013785655610263348, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.013923999853432178, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.013539589010179043, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00953877903521061, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.0943874716758728, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.0943874716758728, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.029126595705747604, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.025182634592056274, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.02395043522119522, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.021196559071540833, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.014604528434574604, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.011888997629284859, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.01754072681069374, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.015519942156970501, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.015203914605081081, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.012454429641366005, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.012199140153825283, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.010500836186110973, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.009078937582671642, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.006682392209768295, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.006567288655787706, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.004498036112636328, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.003805094864219427, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.003626737277954817, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.0033917997498065233, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.0031296268571168184, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.0026953350752592087, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.0027921136934310198, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.002278408734127879, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.0020232168026268482, "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": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.029126595705747604, "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": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.029126595705747604, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.056031689047813416, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.05170603096485138, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.04989607259631157, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.04490255191922188, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.025556696578860283, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.023869313299655914, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.029566559940576553, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.027218211442232132, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.026057245209813118, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.02272011712193489, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.02162761427462101, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.014968051575124264, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.012969124130904675, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.012236187234520912, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.012060676701366901, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.007478802464902401, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.006335531361401081, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.006265044212341309, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.005772281438112259, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.005660915281623602, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.003919701091945171, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.003959295339882374, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0036725939717143774, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0026623390149325132, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.056031689047813416, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.056031689047813416, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.049231838434934616, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.04536091536283493, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.04372147470712662, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.03936472907662392, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.02242574840784073, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.020929284393787384, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.026194652542471886, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.02399001270532608, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.022879978641867638, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.019937623292207718, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.019050603732466698, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.013260149396955967, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.01142571959644556, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.010736131109297276, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.010574134066700935, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.00663116155192256, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.005553155206143856, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.0054838028736412525, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.005058008246123791, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.004952333401888609, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.003471959615126252, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0034663069527596235, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.003239360172301531, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0023083710111677647, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.049231838434934616, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.049231838434934616, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1617102324962616, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.15111833810806274, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1476244330406189, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.13396495580673218, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07485278695821762, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.07110810279846191, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08391123265028, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07743437588214874, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07600996643304825, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06730367243289948, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06399255245923996, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.042527418583631516, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03688696026802063, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.035713233053684235, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.035438187420368195, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.021204272285103798, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.018177352845668793, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01809033751487732, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.016630228608846664, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.016455799341201782, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.010930660180747509, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.010731467977166176, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.010497660376131535, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00670786015689373, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08391123265028, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08391123265028, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.17431002855300903, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.15391112864017487, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.14612261950969696, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1186821311712265, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08139549940824509, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.07355891168117523, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0962158739566803, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08758141100406647, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08362498879432678, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06517191231250763, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05802350863814354, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04981370270252228, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04275405779480934, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.039933882653713226, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03923626244068146, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02534627914428711, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.022006632760167122, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02176816202700138, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.019074147567152977, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.018636321648955345, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014511811546981335, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.015480856411159039, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013644954189658165, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012247801758348942, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0962158739566803, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0962158739566803, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.17112164199352264, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1622736155986786, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.15971392393112183, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.14676114916801453, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07945381104946136, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07652410864830017, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08756310492753983, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08106983453035355, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08032119274139404, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07285794615745544, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06988835334777832, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04441457986831665, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.038654547184705734, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.037934232503175735, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03776714578270912, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.022172527387738228, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01951998472213745, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01947697624564171, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018196815624833107, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01809258759021759, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011635340750217438, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011764353141188622, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011393596418201923, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00786835141479969, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08756310492753983, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08756310492753983, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.2073403298854828, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1967732459306717, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19366255402565002, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.178008571267128, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09626417607069016, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09273470193147659, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10599584877490997, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09814266860485077, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09731237590312958, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0882965624332428, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08470019698143005, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05368449166417122, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04672379791736603, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.0458691380918026, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04566646367311478, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.026761043816804886, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.023304395377635956, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02325265295803547, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.021672304719686508, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.021548038348555565, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.013821358792483807, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.013576728291809559, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.013536994345486164, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008467059582471848, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09814266860485077, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09814266860485077, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.2041136920452118, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18221813440322876, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1743011474609375, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.1550009548664093, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09264408051967621, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08461258560419083, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.10788307338953018, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.09930215775966644, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09570625424385071, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.07991713285446167, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07530568540096283, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05488920584321022, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.04746564105153084, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.044490329921245575, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04376158490777016, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.027499806135892868, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.02325456589460373, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.023129133507609367, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02077716775238514, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.020316535606980324, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.014803915284574032, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.01489727757871151, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.013861186802387238, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.01036842167377472, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09570625424385071, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09570625424385071, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.06558620929718018, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0604901984333992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.05828198790550232, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.05237279459834099, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.029969533905386925, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.027954768389463425, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0347096286714077, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.032086826860904694, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.030562924221158028, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.026594828814268112, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.025232020765542984, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.017570992931723595, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.015309583395719528, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.014377051964402199, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.014156466349959373, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.008794599212706089, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.007496447768062353, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.007402636576443911, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.0068276310339570045, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.006689808331429958, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.004636015277355909, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.00476452149450779, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.00432472862303257, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0032857521437108517, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.06558620929718018, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.06558620929718018, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.0563201978802681, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.051929738372564316, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.04989241436123848, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.044845595955848694, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.025758400559425354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.02396342158317566, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.03023923933506012, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.027744805440306664, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.026265989989042282, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.022863086313009262, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.021824490278959274, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.015314619988203049, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.013227328658103943, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.012358207255601883, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.012143378145992756, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.007665181066840887, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.006413373164832592, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.006325756665319204, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.005839781370013952, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.005706836469471455, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.004030168056488037, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.004055671393871307, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0037394461687654257, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0027429857291281223, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0563201978802681, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0563201978802681, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.17380739748477936, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.16232213377952576, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.15849263966083527, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1435646116733551, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08059942722320557, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.07647419720888138, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09054932743310928, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08346002548933029, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08184445649385452, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07229877263307571, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0685928612947464, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.045929983258247375, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03977183997631073, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03846871852874756, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03817006200551987, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.022913340479135513, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.019595004618167877, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01949470303952694, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.017887530848383904, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.017689982429146767, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.011829703114926815, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.011590566486120224, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.011340681463479996, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.007257623597979546, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09054932743310928, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09054932743310928, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.17075473070144653, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14921291172504425, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.14178213477134705, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11763367056846619, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07938926666975021, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.07115408033132553, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0929052010178566, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08480925858020782, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0817032977938652, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06433434784412384, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05690857023000717, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04793938249349594, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04127642139792442, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.038862619549036026, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03829047083854675, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.024330059066414833, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02129439078271389, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.021114446222782135, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.018534403294324875, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.018168658018112183, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013785562478005886, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014802450314164162, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0130311893299222, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0116371288895607, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0929052010178566, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0929052010178566, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1600128412246704, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15142205357551575, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.14882971346378326, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13637572526931763, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07451906055212021, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07161850482225418, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08235205709934235, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07618027180433273, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07540116459131241, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06805871427059174, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06517589092254639, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04193367063999176, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03644043579697609, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03569700941443443, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03552623838186264, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.020939987152814865, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.018533194437623024, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.018484536558389664, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.017245609313249588, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.017142072319984436, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011127326637506485, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.01141909509897232, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010886996984481812, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007929101586341858, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08235205709934235, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08235205709934235, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21303002536296844, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20166277885437012, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19831162691116333, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1818876564502716, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09921522438526154, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09539290517568588, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10951342433691025, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10137122869491577, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10036812722682953, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0906262993812561, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08670002967119217, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05550666153430939, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04828247055411339, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.047322724014520645, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04709756746888161, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.027694685384631157, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024050800129771233, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.023992737755179405, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02228178083896637, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022137507796287537, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014334325678646564, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014033115468919277, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014004768803715706, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008749354630708694, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09921522438526154, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09921522438526154, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20458626747131348, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.1826198548078537, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.17439499497413635, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.1550748646259308, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09300542622804642, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08476237207651138, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.1089576780796051, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10017167031764984, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09608466923236847, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08027739077806473, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07554212957620621, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.055502183735370636, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.04796494171023369, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04475349187850952, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.043975524604320526, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02788567915558815, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.023538969457149506, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.023385850712656975, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.021072110161185265, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.020581109449267387, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015187550336122513, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.015288800932466984, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014179199934005737, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.010831119492650032, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09608466923236847, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09608466923236847, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08028453588485718, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07488320767879486, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.07269678264856339, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0658176839351654, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.037100762128829956, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03499848395586014, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04220465198159218, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03904072940349579, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.037675000727176666, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0332808643579483, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03161409869790077, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.021406229585409164, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.01865360327064991, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.017790677025914192, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.017583046108484268, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01071776170283556, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009256232529878616, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009174234233796597, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.008499876596033573, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.008372873067855835, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005657801404595375, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.00580934714525938, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005373266525566578, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004010830540210009, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08028453588485718, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08028453588485718, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.06641372293233871, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.06194150820374489, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.06005239859223366, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0544033907353878, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.030697191134095192, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.028927892446517944, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.03540096804499626, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.032492171972990036, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.031187046319246292, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.027592599391937256, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.026385167613625526, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.017969049513339996, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.015511129051446915, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.014715548604726791, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.01452112477272749, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.008988454006612301, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0076081762090325356, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.00752677209675312, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.006984843406826258, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.006861037574708462, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0047097099013626575, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.004726959392428398, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.004443194251507521, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0031642934773117304, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.06641372293233871, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.06641372293233871, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1888355314731598, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17710867524147034, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1733405441045761, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1575833112001419, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0878756195306778, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08369102329015732, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09825682640075684, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09062258154153824, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08912482857704163, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07925274968147278, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07530560344457626, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.049786314368247986, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04317964240908623, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04193638637661934, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.041631657630205154, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02483748458325863, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.021314311772584915, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.021217530593276024, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01953372359275818, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.019342567771673203, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.012769562192261219, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.012501118704676628, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.01229996420443058, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.007735434919595718, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09825682640075684, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09825682640075684, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.16861213743686676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1481700837612152, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.13995923101902008, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11731720715761185, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07841507345438004, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.06991396844387054, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09381741285324097, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08526959270238876, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08049675822257996, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06283332407474518, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.058645881712436676, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04835350438952446, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04155006259679794, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03836992383003235, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.037589576095342636, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02442479133605957, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.020998382940888405, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02070521004498005, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01815180853009224, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.017664099112153053, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01377539336681366, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014708928763866425, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012784882448613644, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011473584920167923, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09381741285324097, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09381741285324097, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.15677496790885925, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14817218482494354, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.14555922150611877, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13329006731510162, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07319378852844238, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07026442140340805, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08090809732675552, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.0749010369181633, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07409751415252686, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06669516116380692, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0637122169137001, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04107373207807541, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03576372563838959, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.034994423389434814, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03481489047408104, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02050255611538887, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.017973003908991814, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.017926223576068878, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.016659483313560486, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016544919461011887, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010714324191212654, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.010800876654684544, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010458133183419704, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007144193165004253, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08090809732675552, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08090809732675552, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21507704257965088, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20343537628650665, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19993582367897034, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.18302948772907257, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10039374232292175, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09637033194303513, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11089977622032166, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10264228284358978, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10159187763929367, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09151183068752289, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08743192255496979, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05626523122191429, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04893019422888756, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.047901853919029236, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04766619950532913, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028081558644771576, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02436172217130661, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02429422177374363, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.0225350521504879, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022381162270903587, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014555459842085838, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.01424516923725605, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014203159138560295, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008902065455913544, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09151183068752289, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09151183068752289, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20728586614131927, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18527553975582123, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.177016019821167, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15750238299369812, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09432779997587204, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.0860457792878151, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11051708459854126, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10166987031698227, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09748510271310806, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08154338598251343, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07687246799468994, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05628981068730354, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.048706281930208206, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04539388045668602, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04459798336029053, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.028260989114642143, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.023881815373897552, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.023716744035482407, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.021406838670372963, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.020896585658192635, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015363001264631748, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.015530956909060478, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.01432843878865242, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011000475846230984, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09748510271310806, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09748510271310806, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.07791686058044434, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07277600467205048, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.07071076333522797, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06399228423833847, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03599350526928902, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03399694338440895, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04101253300905228, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03787380829453468, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03654072433710098, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03232372924685478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.030757414177060127, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02079183980822563, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.018081964924931526, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.017241109162569046, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.017041586339473724, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.010397369042038918, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.008901163004338741, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.008818946778774261, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.008166749030351639, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.008041405119001865, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005448873620480299, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.005494735203683376, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0051651946268975735, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0036661953199654818, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07791686058044434, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07791686058044434, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.06554152816534042, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.06126870959997177, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.059354960918426514, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.05376337096095085, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03030502237379551, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.028552720323204994, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.034883446991443634, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0321270190179348, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.030775852501392365, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.027248872444033623, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.026002636179327965, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.01767592877149582, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.015341632068157196, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.014514493755996227, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.014316601678729057, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.00884011760354042, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.007491643074899912, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.007406710647046566, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.006874872371554375, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.006753073073923588, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.004618794657289982, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.004639306105673313, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.004348481073975563, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.003080235794186592, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.06554152816534042, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.06554152816534042, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.18515005707740784, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17379626631736755, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1700848489999771, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1544744223356247, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08622182905673981, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08214369416236877, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09668701887130737, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08898541331291199, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08745082467794418, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07770407944917679, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07392803579568863, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04906262829899788, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04239996522665024, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04114769399166107, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04084761440753937, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.024473395198583603, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02091965079307556, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.020822498947381973, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.019171852618455887, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.018977144733071327, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.012602818198502064, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.012291411869227886, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.01210110168904066, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.007612815126776695, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09668701887130737, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09668701887130737, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1909492015838623, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.16054610908031464, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.15007635951042175, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.12512367963790894, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08738333731889725, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.07566581666469574, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10442810505628586, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09508586674928665, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09090529382228851, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06768602132797241, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06265224516391754, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05378445237874985, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04652721434831619, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04302504286170006, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04216684773564339, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02737469971179962, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024003496393561363, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.023743342608213425, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.020533639937639236, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.019999373704195023, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015686145052313805, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01726822555065155, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014620831236243248, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01391170360147953, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09508586674928665, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09508586674928665, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.14705058932304382, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.13898523151874542, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1364869624376297, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.12485474348068237, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.06877255439758301, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.0659535825252533, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07614067941904068, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07048417627811432, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.06962595134973526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06263580918312073, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.059817761182785034, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.03867962211370468, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03369978815317154, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.032935239374637604, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03275470808148384, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.019327297806739807, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01700071059167385, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01694616675376892, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.015763677656650543, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.0156505536288023, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010157288052141666, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.010341319255530834, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.009904907085001469, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007000286597758532, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07614067941904068, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07614067941904068, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21221303939819336, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.2006741464138031, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19715243577957153, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1805555671453476, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09937112033367157, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09537102282047272, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10987921059131622, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10168811678886414, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10058274120092392, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09051761776208878, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08657734841108322, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.055938541889190674, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04859120398759842, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.047562647610902786, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04731013998389244, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.027926210314035416, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02451750822365284, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02445296198129654, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022741522639989853, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02258669212460518, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014692586846649647, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014859922230243683, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.01434453297406435, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.010004853829741478, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09051761776208878, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09051761776208878, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.19968761503696442, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.17753389477729797, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.16942554712295532, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15060955286026, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09071293473243713, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08244650065898895, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.10749486088752747, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.09787306189537048, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09380646049976349, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.07822507619857788, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07377531379461288, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.054513733834028244, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.047108836472034454, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04389801248908043, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04311097040772438, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.027634799480438232, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.023536155000329018, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.0233742818236351, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.021175524219870567, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.020688006654381752, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.01547334250062704, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.015839889645576477, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014495963230729103, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.01182561181485653, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.09787306189537048, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.09787306189537048, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08845049142837524, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08270278573036194, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08042817562818527, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07289537787437439, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.040860969573259354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03865659236907959, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04645049571990967, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04285780340433121, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04148152470588684, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.036743029952049255, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.034997835755348206, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.023535966873168945, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.020453371107578278, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.019562147557735443, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.019352223724126816, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.011771165765821934, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010098367929458618, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010009936057031155, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009275343269109726, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009141288697719574, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006167695391923189, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006222654599696398, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0058664302341639996, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004162712022662163, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08845049142837524, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08845049142837524, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.07221280783414841, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.06756165623664856, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.06556101888418198, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.05941495671868324, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03340386226773262, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0315144807100296, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.03828602284193039, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03529328107833862, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03390724956989288, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.030066778883337975, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.028694063425064087, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.019394274801015854, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.016841379925608635, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.015995241701602936, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.015787534415721893, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.009701677598059177, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.008254444226622581, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.008166342042386532, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.007584573235362768, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.007456323131918907, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005081396084278822, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.00509744044393301, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.004806130193173885, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00338574661873281, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07221280783414841, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07221280783414841, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19497443735599518, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18315041065216064, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.17928175628185272, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.16294342279434204, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09079233556985855, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08651834726333618, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10187701880931854, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09364061057567596, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09202634543180466, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08193417638540268, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07812326401472092, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.051733165979385376, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.044610846787691116, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.043324634432792664, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04301465302705765, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02580036222934723, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.022015850991010666, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02191275916993618, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.020194321870803833, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.020000310614705086, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013291161507368088, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01292169839143753, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012759627774357796, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008004664443433285, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09364061057567596, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09364061057567596, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.17967954277992249, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.15886826813220978, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.152234748005867, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.13033781945705414, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08342082798480988, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.07577884942293167, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09582909196615219, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08815144747495651, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08581405133008957, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06939086318016052, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06244578957557678, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04958534985780716, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04310521483421326, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04098344221711159, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04047305881977081, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.025231938809156418, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02260780707001686, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.022469474002718925, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02003512904047966, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01973198540508747, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01450243592262268, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01583249866962433, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013880835846066475, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012653927318751812, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09582909196615219, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09582909196615219, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.14941193163394928, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14115826785564423, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.13860297203063965, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1269010305404663, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07001586258411407, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06714808940887451, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07756604254245758, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07175417989492416, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07086694240570068, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06375262886285782, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06095735728740692, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.03951982781291008, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.034397926181554794, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03361816331744194, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03343496471643448, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.0197652168571949, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.017505116760730743, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.017452633008360863, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.016271328553557396, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01615472324192524, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010522816330194473, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.010887810960412025, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010267128236591816, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007644902914762497, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07756604254245758, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07756604254245758, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21138659119606018, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.19992110133171082, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19637037813663483, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17979936301708221, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09885280579328537, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09486719220876694, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10953167825937271, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.101227767765522, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10004560649394989, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09008309990167618, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08607389032840729, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.055602557957172394, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.048270486295223236, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.047210246324539185, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.046961527317762375, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.027772439643740654, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02405524253845215, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.023978346958756447, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022248385474085808, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022085631266236305, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014470493420958519, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014136983081698418, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014104319736361504, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008923313580453396, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09885280579328537, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09885280579328537, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20868006348609924, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18583081662654877, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.17711994051933289, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15761882066726685, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09512223303318024, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.0863124206662178, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11229624599218369, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10281708091497421, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09837698191404343, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08184000104665756, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07708050310611725, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05711546540260315, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.04947712644934654, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04600708931684494, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04516949877142906, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.028809985145926476, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.024586956948041916, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.02440635859966278, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.022061994299292564, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.02153298631310463, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015919942408800125, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016470201313495636, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014833040535449982, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.012179387733340263, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09837698191404343, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09837698191404343, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09361496567726135, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08761529624462128, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08522038161754608, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07738843560218811, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.043421562761068344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.041094813495874405, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.049261465668678284, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04551857337355614, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.044068533927202225, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03910677134990692, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03725521266460419, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02509276755154133, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.021842824295163155, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.020891854539513588, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.020671334117650986, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012575257569551468, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010979141108691692, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010889931581914425, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010140701197087765, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010002221912145615, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006703539751470089, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007028783671557903, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006391122471541166, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005033442750573158, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09361496567726135, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09361496567726135, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.07969031482934952, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0746484100818634, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.0725368857383728, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06584233790636063, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03696003556251526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03495338559150696, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.042222585529088974, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.038938969373703, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.037505026906728745, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03331531211733818, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03180868178606033, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.021436795592308044, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.018596835434436798, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.017706559970974922, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.017501670867204666, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.0107251713052392, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009150763973593712, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009064026176929474, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.008422290906310081, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.008289871737360954, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005631486885249615, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0056654298678040504, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005337925162166357, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.003797987475991249, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07969031482934952, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.07969031482934952, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19397728145122528, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18231609463691711, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.17845994234085083, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1623685359954834, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.090355783700943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08615350723266602, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10155133157968521, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09324008971452713, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09160075336694717, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08166666328907013, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07786854356527328, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.051523566246032715, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04447333887219429, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04316512122750282, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.042846113443374634, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.025715138763189316, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.021982047706842422, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.021874243393540382, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.020183302462100983, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.019992752000689507, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013302957639098167, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.012972485274076462, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012777132913470268, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008133098483085632, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09324008971452713, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09324008971452713, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19114342331886292, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.16483744978904724, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.15487486124038696, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1258355677127838, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.087632916867733, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.07740160822868347, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10621640831232071, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09693906456232071, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09129147976636887, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0703277662396431, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06272424012422562, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05493070185184479, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04750295728445053, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.043282829225063324, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04225189611315727, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.027894683182239532, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024326803162693977, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.023989127948880196, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02129419706761837, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.020657092332839966, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.016071360558271408, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01776994578540325, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014803942292928696, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014381008222699165, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09693906456232071, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09693906456232071, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1566472053527832, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1480676382780075, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.14541581273078918, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13311971724033356, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07339458912611008, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07038657367229462, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08130612969398499, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07525365799665451, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.0743160992860794, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06684141606092453, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06394274532794952, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04143478348851204, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.036070145666599274, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03522823378443718, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03503657877445221, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.020716030150651932, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01832883059978485, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.018274765461683273, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.017024854198098183, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016907356679439545, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010996825993061066, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011370415799319744, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010720527730882168, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007952926680445671, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08130612969398499, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08130612969398499, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.2149713784456253, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20335853099822998, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19981174170970917, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.18289032578468323, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.1006087064743042, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09652338176965714, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11141337454319, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10298788547515869, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10180042684078217, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09163997322320938, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0875626653432846, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05652524530887604, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.049118299037218094, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.048036206513643265, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04777813330292702, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028218943625688553, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024462949484586716, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024386683478951454, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022617248818278313, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022453922778367996, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014645552262663841, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014369087293744087, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014269014820456505, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009048146195709705, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09163997322320938, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09163997322320938, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.2190292328596115, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.19534561038017273, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.18634995818138123, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.16529642045497894, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.1000761091709137, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.09093032777309418, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11782778054475784, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10808343440294266, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.1034916564822197, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08611538261175156, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.08110067248344421, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.06023828685283661, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05214359983801842, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04852808639407158, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.047648996114730835, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.03037410043179989, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.02612004056572914, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.02592885121703148, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.023500407114624977, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.022955525666475296, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.016846604645252228, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.0177200548350811, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.015691615641117096, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.013354390859603882, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08611538261175156, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08611538261175156, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09981662034988403, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0935034304857254, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09107548743486404, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08254285901784897, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04625426605343819, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.043837908655405045, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05234859511256218, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04837081581354141, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.046920061111450195, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.041651785373687744, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0396265871822834, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.026578737422823906, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02313651517033577, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.022176863625645638, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.021954767405986786, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013296971097588539, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01150807086378336, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011418008245527744, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010595947504043579, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010453845374286175, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006999932695180178, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007166839204728603, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006677377037703991, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004908266011625528, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0935034304857254, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0935034304857254, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.0808504968881607, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07572825253009796, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.07354701310396194, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06668499112129211, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.037426527589559555, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03537161648273468, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.042783014476299286, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03947671130299568, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03797455504536629, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03371124714612961, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.032147299498319626, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0216683316975832, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.01885242946445942, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.01792999915778637, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.017708007246255875, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.010845119133591652, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009262816049158573, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009165970608592033, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.008516118861734867, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0083792544901371, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005687532480806112, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.005727090872824192, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005389817524701357, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0038257299456745386, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0808504968881607, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0808504968881607, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19459646940231323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18273413181304932, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.17864690721035004, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.16227184236049652, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09057527035474777, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08624594658613205, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10162966698408127, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09376415610313416, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0917971134185791, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08176030963659286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07766253501176834, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05152614414691925, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.044715944677591324, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04324277862906456, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.042906198650598526, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02571641281247139, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02205488458275795, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.021927976980805397, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02025008574128151, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.020024126395583153, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013291931711137295, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.013090998865664005, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012756182812154293, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00824784953147173, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09376415610313416, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09376415610313416, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.20489972829818726, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1874377280473709, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.18054397404193878, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15580721199512482, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0961008369922638, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08928608149290085, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11153431236743927, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10228709876537323, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0981413945555687, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08279432356357574, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07506813853979111, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05767931416630745, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04985292628407478, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.047025538980960846, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.046339262276887894, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.029142409563064575, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025676697492599487, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.025407036766409874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02319197915494442, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022764740511775017, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.016372093930840492, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.017750246450304985, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.015514436177909374, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013876871205866337, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.0981413945555687, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.0981413945555687, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.14894923567771912, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1406569629907608, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1380283534526825, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1262151449918747, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.06981571018695831, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06688906997442245, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07741045206785202, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07162400335073471, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07068832963705063, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06347575038671494, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06064308062195778, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.03951796516776085, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03437865525484085, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03355346992611885, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.0333617702126503, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.01975218579173088, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01755904033780098, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.017504118382930756, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.016315527260303497, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016198109835386276, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010563954710960388, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011034109629690647, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010295934043824673, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007878783158957958, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07741045206785202, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07741045206785202, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.2112918496131897, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.19953438639640808, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19592827558517456, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1791447401046753, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0987454503774643, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09465289860963821, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10940473526716232, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10113847255706787, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09996093809604645, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08979193866252899, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08568640798330307, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.0554656945168972, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04823283106088638, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04715068265795708, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04688849672675133, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.027698025107383728, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02400643564760685, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.023933645337820053, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022162320092320442, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02199787087738514, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014373613521456718, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014093900099396706, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.01400173269212246, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008854844607412815, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0987454503774643, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0987454503774643, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.22019745409488678, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.19744957983493805, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.18918350338935852, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.16842208802700043, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.10075241327285767, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.09208831191062927, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11713773757219315, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10788615047931671, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10386376827955246, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.0872049406170845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.08196978271007538, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05983304977416992, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05193158984184265, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04865516349673271, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.047873612493276596, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.030171044170856476, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.025890672579407692, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.025731243193149567, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02330058254301548, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.02281235344707966, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.016680635511875153, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.01715274341404438, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.015701264142990112, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.012617070227861404, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.0872049406170845, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.0872049406170845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09771290421485901, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09148745983839035, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08905331045389175, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08074796199798584, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04528608173131943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04289918765425682, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.051327675580978394, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04744406417012215, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04594210535287857, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.040765125304460526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03878991678357124, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.026045463979244232, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022653082385659218, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02168739214539528, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.021458249539136887, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013017796911299229, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011203153058886528, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011104976758360863, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010298243723809719, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01015474647283554, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006823237519711256, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0069081843830645084, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006494496483355761, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004631890915334225, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09771290421485901, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09771290421485901, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08075544983148575, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07570365071296692, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.07347812503576279, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06665561348199844, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.037446215748786926, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03539678454399109, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.042757052928209305, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.039553575217723846, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03798789903521538, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03374991565942764, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.032166093587875366, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.021670255810022354, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.01886001043021679, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.017930909991264343, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.01770670711994171, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.010836257599294186, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009237714111804962, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009143666364252567, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.008494670502841473, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.008353057317435741, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005676255561411381, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.005680801346898079, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005377343390136957, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.003751714015379548, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08075544983148575, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08075544983148575, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19603869318962097, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18409480154514313, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.17991149425506592, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.16334357857704163, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09130316227674484, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08688601851463318, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10267359763383865, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09458442777395248, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09255512803792953, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08233213424682617, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07824195921421051, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0520227886736393, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04509695991873741, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.043614860624074936, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04326031357049942, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02597798779606819, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02224181964993477, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.022104674950242043, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02040247432887554, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.020175369456410408, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013427185826003551, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.013202190399169922, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012873874977231026, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008323634043335915, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09458442777395248, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09458442777395248, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19740426540374756, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17644169926643372, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1670297235250473, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.14359168708324432, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0910053551197052, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08264756202697754, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11028867959976196, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1010327935218811, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09460178762674332, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0777261033654213, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07180911302566528, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05705767869949341, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.049114517867565155, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.044427789747714996, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.043269407004117966, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02860875055193901, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024114150553941727, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.023703105747699738, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02163592539727688, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.020922208204865456, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015720149502158165, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.016822518780827522, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014232255518436432, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012762857601046562, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.09460178762674332, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.09460178762674332, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.14966793358325958, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14122943580150604, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.13852760195732117, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.12664802372455597, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07013238966464996, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06714798510074615, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07787460833787918, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07205379009246826, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07102955877780914, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06376305222511292, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06092247739434242, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.03977800905704498, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03461793065071106, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03376677632331848, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03355775773525238, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.01990656740963459, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.017728693783283234, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01766534522175789, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01647549867630005, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01635252684354782, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010685255751013756, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011232399381697178, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010406539775431156, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008102922700345516, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07787460833787918, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07787460833787918, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20901544392108917, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.19725601375102997, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1936405599117279, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1769046038389206, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0976129099726677, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09352555871009827, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10825186967849731, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.1001063734292984, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.0988427996635437, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08872438222169876, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0846024751663208, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05489054694771767, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.047748636454343796, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.046625927090644836, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.046367302536964417, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02739235945045948, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02377106435596943, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02369154617190361, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02194051630795002, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02177351526916027, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014181869104504585, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014003111980855465, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.01379397977143526, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008854398503899574, "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": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.0988427996635437, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0976129099726677, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.22392258048057556, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.2014780193567276, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.19314099848270416, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.1714264452457428, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.10262726247310638, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.09407346695661545, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.1193723976612091, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10987712442874908, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10576413571834564, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08888418972492218, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.08352852612733841, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.06109710782766342, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.052834801375865936, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.049516480416059494, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04871690273284912, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.03081093356013298, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.02625703625380993, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.026096677407622337, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02362273447215557, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.023122774437069893, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.01701405458152294, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.017291024327278137, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.016002006828784943, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.012597120366990566, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08888418972492218, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08888418972492218, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10396752506494522, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09742747247219086, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09486386924982071, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08600371330976486, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04824040085077286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04572286829352379, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05461106076836586, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0504230298101902, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04893122613430023, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04342031851410866, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04127829149365425, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.027659926563501358, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02406350150704384, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.023091981187462807, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02285786345601082, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013834036886692047, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011877821758389473, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011783563531935215, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01090440433472395, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010755512863397598, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007222585845738649, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007243719883263111, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0068904985673725605, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004763912409543991, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09742747247219086, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09742747247219086, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08462570607662201, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07932659983634949, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.07709166407585144, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06992492079734802, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03925636038184166, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0371180921792984, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.044759925454854965, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04138403758406639, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03982200101017952, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03538224846124649, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03371543809771538, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0226681400090456, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.019739443436264992, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.018782934173941612, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.01855749823153019, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.011331862770020962, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009664507582783699, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009564804844558239, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.00888197124004364, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.008737987838685513, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005921171046793461, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.005921011790633202, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005618289113044739, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0038807052187621593, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08462570607662201, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08462570607662201, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2020137906074524, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18961000442504883, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.18524551391601562, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.16806499660015106, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09415072202682495, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08952002972364426, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10600978136062622, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0976206362247467, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09539655596017838, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08481819182634354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08060580492019653, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05374514311552048, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04660845920443535, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04499021917581558, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04460033029317856, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.026870671659708023, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02301873452961445, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.022869743406772614, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02111699804663658, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.020871669054031372, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01397511176764965, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.013772434554994106, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013399305753409863, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008821754716336727, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0976206362247467, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0976206362247467, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.20959386229515076, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1902122050523758, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.18365661799907684, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1605204939842224, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09788066148757935, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09076428413391113, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.1128884106874466, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1032305508852005, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10014234483242035, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08405500650405884, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0779414102435112, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0580417700111866, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05009978264570236, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04763200506567955, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.047043487429618835, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.029207684099674225, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0257419403642416, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02555161900818348, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.0231271144002676, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022770684212446213, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.016254287213087082, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01741190440952778, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.015508218668401241, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013373550958931446, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09788066148757935, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09788066148757935, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.15322788059711456, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14460329711437225, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.14181995391845703, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.12959052622318268, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0718904435634613, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06882057338953018, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07989327609539032, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07386941462755203, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07280278205871582, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06530562788248062, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0623895525932312, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04078967496752739, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.035507507622241974, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.0346171110868454, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03441851586103439, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.020415816456079483, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.018226005136966705, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.018158573657274246, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01693710871040821, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.0168120376765728, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010985546745359898, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011604276485741138, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010698951780796051, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00843470823019743, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07989327609539032, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07989327609539032, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.2114408016204834, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1996341347694397, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19585058093070984, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17886844277381897, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09887496381998062, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09470845758914948, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10975081473588943, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10141738504171371, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10010762512683868, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08979993313550949, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0856560617685318, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05572914332151413, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04837041348218918, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04723373427987099, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04695994406938553, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.027816612273454666, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02404399774968624, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.023960264399647713, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022175388410687447, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02200477197766304, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014431209303438663, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014129672199487686, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014026924036443233, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008879079483449459, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09887496381998062, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09887496381998062, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.22629450261592865, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.20330485701560974, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.19485314190387726, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.17345531284809113, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.10377230495214462, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.0950675904750824, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.12082856893539429, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.11098387837409973, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10700183361768723, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08988770842552185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.08464708924293518, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.06188852712512016, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.053405776619911194, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.050154224038124084, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04936769977211952, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.031303100287914276, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.0267111137509346, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.026553679257631302, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.024058030918240547, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.023569755256175995, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.017478283494710922, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.017705867066979408, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.016471125185489655, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.013051293790340424, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08988770842552185, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08988770842552185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11106713116168976, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10403386503458023, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10140809416770935, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0919741615653038, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05156863108277321, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.048913680016994476, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05826105177402496, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05382252484560013, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.052306097000837326, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.046483080834150314, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.044215522706508636, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02960488572716713, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025744888931512833, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02472347393631935, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.024480564519762993, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01480442751199007, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012820370495319366, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012726221233606339, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011804766952991486, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011655737645924091, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007794691249728203, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007955123670399189, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007456641644239426, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0054346052929759026, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.0919741615653038, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.0919741615653038, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08735247701406479, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0818939059972763, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.07940873503684998, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07205361127853394, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04045283794403076, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03822357952594757, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04632315784692764, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04283810034394264, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04106316715478897, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0364748053252697, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.034786857664585114, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02347460202872753, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02043578028678894, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.019378557801246643, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.019129376858472824, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.011740178801119328, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009990516118705273, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009877586737275124, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009187779389321804, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009026290848851204, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006142154801636934, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006161308381706476, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005807137116789818, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004064424429088831, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08735247701406479, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08735247701406479, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2099318951368332, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19709046185016632, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.19276097416877747, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1749868243932724, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09782720357179642, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09317543357610703, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10999223589897156, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10127370804548264, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09920148551464081, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08830302953720093, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08399951457977295, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05585850775241852, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.048361632972955704, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04677365720272064, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04640460014343262, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.027903592213988304, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02393324114382267, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.023796936497092247, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.021980494260787964, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.021732695400714874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014500271528959274, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014320751652121544, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013911290094256401, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009186400100588799, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.09920148551464081, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09782720357179642, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.20982669293880463, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18776732683181763, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.18073710799217224, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15554462373256683, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09858349710702896, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09044145047664642, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.1134287565946579, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10322577506303787, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10092214494943619, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08247088640928268, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07641258835792542, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.058891329914331436, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.050336889922618866, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.048219695687294006, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.047705505043268204, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02974674291908741, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.026131076738238335, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.025993265211582184, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02309519238770008, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022784722968935966, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.016697673127055168, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01771712675690651, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0160724725574255, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013727951794862747, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09858349710702896, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09858349710702896, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1527356058359146, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14408740401268005, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1412891298532486, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1290675848722458, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07169130444526672, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06860167533159256, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07957392185926437, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07365725934505463, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07261737436056137, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06513245403766632, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.062106844037771225, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.040615446865558624, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.035402681678533554, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03452962636947632, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03432008996605873, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.020327655598521233, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.018129417672753334, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.018072612583637238, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.016846777871251106, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016717225313186646, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010897104628384113, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011499634943902493, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010610884055495262, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008308430202305317, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07957392185926437, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07957392185926437, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.212095245718956, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.2001720666885376, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19646425545215607, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17943470180034637, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09924829751253128, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09509244561195374, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10998974740505219, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10171567648649216, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10050369799137115, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09015157073736191, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08589980006217957, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05582861974835396, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04852283373475075, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.047398053109645844, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04713423177599907, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02787601202726364, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02412903495132923, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024055158719420433, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022254154086112976, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022082922980189323, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014463870786130428, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014161366038024426, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014068622142076492, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00888817198574543, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09924829751253128, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09924829751253128, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.22757117450237274, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.20390567183494568, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.19512850046157837, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.17374533414840698, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.10396371781826019, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.09501338750123978, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.12202833592891693, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.11165456473827362, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10737056285142899, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08999447524547577, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.08481200039386749, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.06212807074189186, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05358891561627388, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.050146907567977905, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04930504783987999, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.031203284859657288, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.02652846835553646, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.026367444545030594, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.023837396875023842, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.02331474795937538, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.017042070627212524, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.017413359135389328, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.0159433763474226, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.012584607116878033, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08999447524547577, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08999447524547577, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11326266080141068, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10629904270172119, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10365720093250275, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09417116641998291, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05264214798808098, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.049998871982097626, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05939261242747307, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05486149340867996, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05335439369082451, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0475151389837265, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0452263168990612, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03013048693537712, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.026201117783784866, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.025185126811265945, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.024945218116044998, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.015055808238685131, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012964884750545025, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01286755595356226, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011929566040635109, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011779505759477615, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007860790006816387, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007899698801338673, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007514630444347858, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005209571216255426, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09417116641998291, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09417116641998291, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08790085464715958, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0825156569480896, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08016645163297653, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07283082604408264, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.040806129574775696, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.038613442331552505, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04654742404818535, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.043059784919023514, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04138898476958275, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03686320036649704, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.035122230648994446, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.023570263758301735, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.020529942587018013, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.019520040601491928, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.019281992688775063, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.011772803962230682, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010023759678006172, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009918555617332458, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009223353117704391, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009072631597518921, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006135637406259775, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006118185352534056, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005814336705952883, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.003973115235567093, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08790085464715958, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08790085464715958, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21373240649700165, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2010839581489563, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.19648388028144836, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.17884787917137146, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09992384165525436, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0951097160577774, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11255018413066864, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10360461473464966, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10127579420804977, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09033871442079544, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08603835105895996, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057116635143756866, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.049544550478458405, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04778686538338661, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04736285284161568, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028544196859002113, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024380434304475784, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.024227604269981384, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022419149056077003, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022153610363602638, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014833273366093636, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01453139167279005, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014209588058292866, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009202116169035435, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09033871442079544, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09033871442079544, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2086879163980484, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18743659555912018, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1807613968849182, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15734083950519562, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09708111733198166, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08960571885108948, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11093743145465851, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10128837823867798, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09920787811279297, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08194392174482346, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0764160230755806, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057156119495630264, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04945473372936249, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04758257791399956, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04714549705386162, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02919292077422142, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02613014727830887, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.0260095726698637, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02332533895969391, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02305642142891884, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.016871171072125435, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.018068527802824974, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.016339631751179695, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014381911605596542, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.09920787811279297, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.09920787811279297, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.15353526175022125, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14489923417568207, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.14211204648017883, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.12976834177970886, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07205306738615036, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06898820400238037, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07998187094926834, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07401362806558609, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07296924293041229, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06544288247823715, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06245633587241173, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.040826354175806046, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.0355190671980381, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03464953228831291, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03443446755409241, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.020414622500538826, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.018116246908903122, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01805964671075344, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.016814639791846275, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016688266769051552, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010910377837717533, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011378934606909752, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010627713054418564, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008100702427327633, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07998187094926834, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07998187094926834, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20962177217006683, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.19789308309555054, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19425012171268463, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17748688161373138, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09811924397945404, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09401356428861618, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10868535935878754, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10056981444358826, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09936029464006424, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08915343135595322, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08499995619058609, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05515998974442482, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04798205941915512, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04686634615063667, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04660111665725708, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02752874791622162, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.023862002417445183, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02378082647919655, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022011736407876015, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.021842531859874725, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.01424243114888668, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014014295302331448, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.013859464786946774, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008809098973870277, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09811924397945404, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09811924397945404, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.2235109955072403, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.19907543063163757, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.18993212282657623, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.1685279756784439, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.10195610672235489, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.09265059232711792, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.1198648139834404, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.11000114679336548, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10547760874032974, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08762798458337784, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.08236462622880936, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.06108144670724869, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05276016145944595, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.049103744328022, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04821169376373291, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.030795549973845482, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.02585180290043354, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.025669384747743607, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.023075439035892487, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.022505201399326324, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.016879767179489136, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016846703365445137, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.015742512419819832, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011951754800975323, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08762798458337784, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08762798458337784, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10822311043739319, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10170215368270874, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09919985383749008, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09021633863449097, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0503469742834568, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04785187169909477, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05682366341352463, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05248720198869705, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05103259161114693, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04549599811434746, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04341597482562065, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02884148806333542, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025062911212444305, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02409346029162407, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.023861102759838104, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014422564767301083, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012377344071865082, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012289800681173801, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011393179185688496, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01125024352222681, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007527240086346865, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007509730756282806, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007204622961580753, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00491687236353755, "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": 2.62529296875, "total_bits": 68820480.0, "err": 0.09919985383749008, "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": 2.62529296875, "total_bits": 68820480.0, "err": 0.09919985383749008, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.0871429592370987, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08181697130203247, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.07954537123441696, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07231874018907547, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.040456049144268036, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03830353915691376, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04605337604880333, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.042633309960365295, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.041010983288288116, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0365527868270874, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.034863393753767014, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.023320525884628296, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02033507637679577, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.019360123202204704, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.01912502758204937, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.011661157943308353, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009951373562216759, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009850146248936653, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009165751747786999, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009012931026518345, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006088145077228546, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006086545996367931, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005781460087746382, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.003972937818616629, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0871429592370987, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0871429592370987, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21687790751457214, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20419606566429138, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.19972699880599976, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18189357221126556, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10140639543533325, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0966690331697464, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.1138901486992836, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10488149523735046, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10271650552749634, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09174570441246033, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08744014799594879, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057811837643384933, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05009530112147331, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04847085475921631, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0480768196284771, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028893275186419487, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024748768657445908, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02460397034883499, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022771235555410385, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022523092105984688, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015009202063083649, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014724389649927616, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014410646632313728, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009354096837341785, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09174570441246033, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09174570441246033, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19854643940925598, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17933295667171478, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1711767613887787, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15171843767166138, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0928109884262085, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08423492312431335, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11015917360782623, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10051079839468002, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09502943605184555, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07993055880069733, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07447549700737, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.056585703045129776, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04848158732056618, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04498065263032913, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.044129494577646255, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028368201106786728, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.023857172578573227, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02350831776857376, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.021352261304855347, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.020813465118408203, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015238397754728794, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.015793707221746445, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014070571400225163, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011432266794145107, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.09502943605184555, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.09502943605184555, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.15175704658031464, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14324791729450226, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1405496746301651, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.12839405238628387, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07123406231403351, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06822215765714645, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07900400459766388, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07312272489070892, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07212663441896439, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0647272914648056, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06176723539829254, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.040302108973264694, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03513684868812561, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.034286078065633774, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03409010171890259, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.020181572064757347, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01799418032169342, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01793689653277397, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.016715774312615395, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01659001037478447, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010815918445587158, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011388720013201237, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010541248135268688, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008205059915781021, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07900400459766388, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07900400459766388, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20808176696300507, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1965395212173462, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19295312464237213, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17638328671455383, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0974198654294014, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.0933508425951004, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10794993489980698, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09984409809112549, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09862115979194641, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08853248506784439, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08446961641311646, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.054772719740867615, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04763060435652733, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04653851315379143, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04628056660294533, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.027347592636942863, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02372012473642826, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.023645659908652306, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02189633809030056, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.021735914051532745, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014202568680047989, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.013968686573207378, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.01383040938526392, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008843851275742054, "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": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09862115979194641, "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": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09862115979194641, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.2222791612148285, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.1981046497821808, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1889180690050125, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.1678011417388916, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.10142477601766586, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.09214501082897186, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11921995878219604, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10965888202190399, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10500789433717728, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08729319274425507, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.08232173323631287, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.060931071639060974, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05263321101665497, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.048876211047172546, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.047969620674848557, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.03073418326675892, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.025782715529203415, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.02558952197432518, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.023056166246533394, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.022470708936452866, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.01686731167137623, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.01688208617269993, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.01570960134267807, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.012035567313432693, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08729319274425507, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08729319274425507, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11442386358976364, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10758950561285019, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10506173968315125, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09569436311721802, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05333159118890762, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0507715567946434, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06004704535007477, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05543728917837143, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.054040465503931046, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04827374219894409, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04608742520213127, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.030522089451551437, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.026507366448640823, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.025560030713677406, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.025331716984510422, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.015273875556886196, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013207235373556614, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01312047615647316, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.012190446257591248, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012048955075442791, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.008021390065550804, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.008113503456115723, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007705427706241608, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005475756712257862, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09569436311721802, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09569436311721802, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.0911678746342659, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08573686331510544, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08345001935958862, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07589445263147354, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.042399872094392776, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0402165912091732, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.048127446323633194, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04454786702990532, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04296578839421272, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.038349296897649765, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03658280521631241, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.024375317618250847, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.021250270307064056, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.020287353545427322, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02006136253476143, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012185400351881981, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01043619029223919, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010339469648897648, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009622721932828426, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009477141313254833, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006361612118780613, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006380257196724415, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006060045212507248, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004185759462416172, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0911678746342659, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0911678746342659, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.22465623915195465, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2118590623140335, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.207464337348938, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18926747143268585, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10519976913928986, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.10048270970582962, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11777547001838684, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10854886472225189, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10655142366886139, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.095370352268219, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09093575924634933, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05980199575424194, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.051839493215084076, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.05026601254940033, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04989113658666611, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.029841123148798943, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02560788206756115, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.025476116687059402, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02359636314213276, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.023352010175585747, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015417286194860935, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01514134556055069, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014840184710919857, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009514127857983112, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.095370352268219, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.095370352268219, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.20104146003723145, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17821793258190155, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.17072394490242004, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.14570845663547516, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09275095164775848, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08437110483646393, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.1076056957244873, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09812334179878235, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09541730582714081, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07611098140478134, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06933912634849548, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05538536608219147, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.047871895134449005, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04545315355062485, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.044862933456897736, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028247982263565063, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024972224608063698, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.024801379069685936, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.021936139091849327, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02155742421746254, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.016116298735141754, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.017394334077835083, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.015394757501780987, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013781338930130005, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09812334179878235, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09812334179878235, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.14877843856811523, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14049910008907318, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.13786934316158295, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1261153370141983, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.06980008631944656, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06688325107097626, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07747451961040497, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07159359753131866, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07066623121500015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06348539143800735, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06070414558053017, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.039470165967941284, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.034363988786935806, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.033548176288604736, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.033359333872795105, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.019767044112086296, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01753854565322399, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01748528890311718, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.016294894739985466, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016171913594007492, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010566267184913158, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011005125008523464, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010301683098077774, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007823551073670387, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07747451961040497, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07747451961040497, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20460140705108643, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1933143436908722, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.18978965282440186, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17363685369491577, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09572791308164597, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09177246689796448, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.1060560792684555, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09806175529956818, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09689036011695862, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08705049008131027, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08322572708129883, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05384719371795654, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04677614942193031, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04572524130344391, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04547858610749245, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.026929713785648346, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.023319445550441742, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02324695698916912, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02154042012989521, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.021384259685873985, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014066561125218868, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.013737629167735577, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.013716006651520729, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008711574599146843, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09806175529956818, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09806175529956818, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.21362997591495514, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18921779096126556, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.17951026558876038, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15936049818992615, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09720069915056229, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08755229413509369, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11550512909889221, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10608676075935364, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.1008351594209671, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08318012207746506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07846692949533463, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.0590389184653759, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05088000372052193, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04685668274760246, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04587576538324356, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02979026362299919, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.024731364101171494, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.024505026638507843, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.022011233493685722, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.021384432911872864, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.016352785751223564, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016270536929368973, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.015087053179740906, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011560075916349888, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09720069915056229, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09720069915056229, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11147022247314453, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10498335212469101, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10258207470178604, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09351181983947754, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.051888808608055115, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.049456410109996796, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05848921462893486, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05397792533040047, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05256125330924988, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.047084685415029526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.045047327876091, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.029710758477449417, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025780711323022842, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.024828175082802773, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02460561692714691, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01484835147857666, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012760112062096596, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012671410106122494, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011792121455073357, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011651928536593914, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0077536385506391525, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007743693422526121, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007434062659740448, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005081846844404936, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09351181983947754, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09351181983947754, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08975152671337128, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08445049822330475, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08222272247076035, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0748853012919426, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.041694268584251404, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03957325965166092, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.047421809285879135, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04384700208902359, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04224569723010063, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03783021122217178, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.036166612058877945, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02402060478925705, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.020926859229803085, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.019957266747951508, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.019729802384972572, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012007523328065872, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010269894264638424, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010168468579649925, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009485971182584763, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009337383322417736, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.00627725338563323, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006282673683017492, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005972261540591717, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004123816266655922, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08975152671337128, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08975152671337128, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2169008105993271, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20474323630332947, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20077352225780487, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1833295077085495, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10145530104637146, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09705884754657745, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.1133882999420166, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10455314815044403, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10271374881267548, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09221439063549042, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08809855580329895, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057514697313308716, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.049892351031303406, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04846120998263359, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04812539368867874, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02872232161462307, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02467939257621765, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.024560710415244102, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02276669628918171, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022555960342288017, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014839447103440762, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014546755701303482, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014301998540759087, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009122777730226517, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09221439063549042, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09221439063549042, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19814731180667877, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17989562451839447, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.17376026511192322, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1441008299589157, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09287214279174805, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08629230409860611, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.106361985206604, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0975014790892601, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09493280202150345, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07758548855781555, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06841465085744858, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0547957569360733, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04732736945152283, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04522128403186798, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0447273887693882, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.027552543208003044, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02436860464513302, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02421717531979084, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02148389257490635, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.021173492074012756, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01519998349249363, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01638207398355007, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014557115733623505, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012532454915344715, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0975014790892601, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0975014790892601, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1519942283630371, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.143720343708992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.14108139276504517, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.12913067638874054, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07135162502527237, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06839411705732346, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07902859896421432, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.0731101781129837, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07220937311649323, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06496886163949966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0620901957154274, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04025667905807495, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.035070061683654785, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03425644338130951, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.034072890877723694, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.020178616046905518, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.017852328717708588, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.017797119915485382, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01658371277153492, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016468344256281853, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010803834535181522, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011104777455329895, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010539660230278969, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007801396772265434, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07902859896421432, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07902859896421432, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20991384983062744, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.19850508868694305, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19500009715557098, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17856694757938385, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09819571673870087, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09423327445983887, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10891298949718475, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10053948312997818, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09940572828054428, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08945733308792114, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08550989627838135, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.055252522230148315, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04797451198101044, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.046924252063035965, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04666844382882118, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02766876481473446, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.023960746824741364, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02387901023030281, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02215670980513096, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022000296041369438, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014557251706719398, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014146793633699417, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014214936643838882, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009026778861880302, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09819571673870087, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09819571673870087, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.21666796505451202, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.19145998358726501, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.18163572251796722, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.16092011332511902, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09861021488904953, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.0886647030711174, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11663997918367386, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10748062282800674, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10226774960756302, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08406408876180649, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.0792631208896637, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05951147899031639, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05172356590628624, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04762447625398636, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04661116003990173, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.0299795214086771, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.025255385786294937, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.025018727406859398, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.022455742582678795, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.021816693246364594, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.016149437054991722, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016770659014582634, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014825236983597279, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.012096058577299118, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09861021488904953, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09861021488904953, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11174529045820236, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1054239422082901, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10302060842514038, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09401696175336838, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05205637961626053, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04966079816222191, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05869526043534279, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05410126596689224, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05270662531256676, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.047317199409008026, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04531353712081909, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.029784558340907097, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025843745097517967, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.024912983179092407, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02469584345817566, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014893575571477413, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012810162268579006, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01272158045321703, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011851209215819836, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011710703372955322, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007793570403009653, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007781710475683212, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007491457741707563, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005132014397531748, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09401696175336838, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09401696175336838, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09257452934980392, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08726581186056137, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08508539944887161, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07761429995298386, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.043070558458566666, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04097061976790428, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04880465939640999, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04509715363383293, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.043627385050058365, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03912098705768585, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03742721676826477, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02472742274403572, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.021517619490623474, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.0206013023853302, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.020381780341267586, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012354947626590729, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010574760846793652, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010481877252459526, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009773734025657177, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.00963601190596819, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0064429184421896935, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0064187925308942795, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006151639856398106, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004172988701611757, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09257452934980392, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09257452934980392, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21898946166038513, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2069065123796463, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20290599763393402, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1855529099702835, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10241597890853882, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09805653244256973, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11419881135225296, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10526265949010849, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10365854948759079, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09312980622053146, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08900794386863708, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057946592569351196, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05020317807793617, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.048878878355026245, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.048559315502643585, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028892559930682182, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024817347526550293, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02470802515745163, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022901296615600586, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02270067110657692, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014869368635118008, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01450507901608944, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014365900307893753, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008943244814872742, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09312980622053146, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09312980622053146, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.212596595287323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1832750141620636, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.17384038865566254, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.14754042029380798, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09811306744813919, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08666739612817764, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.1142820194363594, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10482792556285858, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10173603147268295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08000264316797256, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07135874032974243, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.059003718197345734, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05121282488107681, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04822644963860512, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04749590530991554, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.03008415363729, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.026661522686481476, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.026489464566111565, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.023370612412691116, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022927774116396904, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01742788776755333, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.018820088356733322, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.01656208001077175, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.015048245899379253, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09811306744813919, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09811306744813919, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.15918178856372833, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1505424827337265, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.14786043763160706, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13538013398647308, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0746629536151886, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07165408879518509, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08255670219659805, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07647893577814102, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07556655257940292, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06803432106971741, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06504011899232864, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04203425720334053, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03663502633571625, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03580725938081741, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03561628609895706, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.021002721041440964, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01857719011604786, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.018521172925829887, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.017251836135983467, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.017133299261331558, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.01108916662633419, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011442557908594608, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010823034681379795, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007906079292297363, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08255670219659805, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08255670219659805, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21443584561347961, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20284096896648407, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1992679238319397, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.18251082301139832, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10032552480697632, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09631580114364624, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11082212626934052, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10266119241714478, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10153599083423615, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0914434939622879, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08735465258359909, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.0562373548746109, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.0489480122923851, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.0478975810110569, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.047652699053287506, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028042467311024666, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024372752755880356, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02429724484682083, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022540684789419174, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02237819880247116, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014483117498457432, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.01426857803016901, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.01412822213023901, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008939534425735474, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0914434939622879, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0914434939622879, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.21964333951473236, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.19324593245983124, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.18269897997379303, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.16174718737602234, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09964200109243393, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08926426619291306, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11861360818147659, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10932508856058121, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10360647737979889, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.0847182422876358, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07992084324359894, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.060401204973459244, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05239984765648842, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04802367091178894, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04695240780711174, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.030306778848171234, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.0253314021974802, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.025085197761654854, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02243811823427677, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.021748315542936325, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.016294511035084724, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016683628782629967, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.01490282453596592, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.01179750170558691, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.0847182422876358, "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": 3.523441116898148, "total_bits": 249385216.0, "err": 0.0847182422876358, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11042942106723785, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10409210622310638, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10167737305164337, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09285390377044678, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05142083764076233, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0490017794072628, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05806531012058258, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05352724716067314, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05207517370581627, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04670437052845955, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04475558176636696, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02949107252061367, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025596858933568, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.024628113955259323, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02440604940056801, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01476125419139862, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012705095112323761, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012615418992936611, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011764546856284142, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011622314341366291, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.00773870712146163, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007781934924423695, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0074286628514528275, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0051993075758218765, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09285390377044678, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09285390377044678, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09253958612680435, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08717384934425354, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08481811732053757, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.077445387840271, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04304526746273041, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.040874309837818146, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04893691465258598, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04524261876940727, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0435832254588604, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03909150883555412, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.037397000938653946, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.024795500561594963, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.021581100299954414, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.020587032660841942, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.020349597558379173, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012397552840411663, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010571002960205078, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010469192638993263, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009772107936441898, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009622731246054173, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0064689298160374165, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006438479293137789, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006158723961561918, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004182676784694195, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09253958612680435, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09253958612680435, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21899296343326569, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2069336622953415, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20288068056106567, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18539203703403473, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10238590836524963, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09801105409860611, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11424834281206131, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10543570667505264, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10364712029695511, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0931272879242897, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08896541595458984, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057975634932518005, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05025274306535721, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.048874903470277786, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.048555877059698105, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028915077447891235, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024844903498888016, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.024735858663916588, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02294229343533516, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022727638483047485, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014890264719724655, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014590383507311344, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014369022101163864, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009078803472220898, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.0931272879242897, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.0931272879242897, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19606535136699677, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17186690866947174, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.16415782272815704, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1368032693862915, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0910487174987793, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08180557936429977, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10510098934173584, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09602772444486618, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09360174089670181, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07368230074644089, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06469912827014923, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0541178397834301, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04650044068694115, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.044279541820287704, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0437663309276104, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.027395322918891907, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02383066713809967, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.023685812950134277, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.020534241572022438, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.020196793600916862, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015399254858493805, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.015981659293174744, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014748089015483856, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012164299376308918, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09602772444486618, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09602772444486618, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.16174961626529694, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15309089422225952, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.15032632648944855, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13784702122211456, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07584720849990845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07278572767972946, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08409228920936584, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07766197621822357, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07675904035568237, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0691455528140068, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06624702364206314, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04278359189629555, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03722899407148361, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03640677034854889, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03621164709329605, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.021459471434354782, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.018971025943756104, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.018915917724370956, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.017636453732848167, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.017520006746053696, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011513715609908104, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011793224141001701, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011247900314629078, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008285203017294407, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08409228920936584, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08409228920936584, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.2126682698726654, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20134398341178894, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19784438610076904, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1813068836927414, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09951510280370712, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09557269513607025, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11007566004991531, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10181419551372528, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10072550922632217, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09073755890130997, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08679959923028946, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05586806312203407, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.048609960824251175, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04756104201078415, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04731970280408859, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02795438840985298, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024261053651571274, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02419363334774971, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02246418036520481, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022310050204396248, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014637911692261696, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.01430425327271223, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014300712384283543, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009104489348828793, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09073755890130997, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09073755890130997, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.21542970836162567, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18869549036026, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1776321679353714, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15718959271907806, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09762296825647354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08681721240282059, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11705959588289261, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10790329426527023, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10165739804506302, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08268045634031296, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.0780571773648262, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05962337553501129, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05166539549827576, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04702478647232056, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04587705805897713, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.029872210696339607, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.024720218032598495, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.024438606575131416, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.021809682250022888, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.021061599254608154, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.01601102389395237, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016247009858489037, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014544792473316193, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.01132273767143488, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09762296825647354, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09762296825647354, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10680729150772095, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10077813267707825, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09837789833545685, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08983192592859268, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04969923570752144, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.047366201877593994, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05614040419459343, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.051811277866363525, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05033336579799652, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04520365968346596, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.043283335864543915, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02846216782927513, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.024732613936066628, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02377249300479889, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.023542866110801697, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014219844713807106, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012208128347992897, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012116555124521255, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01129850558936596, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011156504973769188, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007431385572999716, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007407676428556442, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007119161542505026, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0048387120477855206, "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": 2.62529296875, "total_bits": 68820480.0, "err": 0.09837789833545685, "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": 2.62529296875, "total_bits": 68820480.0, "err": 0.09837789833545685, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08997026830911636, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0848044902086258, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08253731578588486, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07534155249595642, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04180076718330383, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03969268873333931, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04760473221540451, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.043967120349407196, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.042333345860242844, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03801017254590988, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.036381375044584274, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.024105338379740715, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.020982632413506508, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.019993094727396965, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.01976338028907776, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012045470997691154, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010273301042616367, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010173171758651733, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009499474428594112, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009354834444820881, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006288438104093075, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.00626535527408123, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0059777554124593735, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004079731181263924, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08997026830911636, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08997026830911636, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21130353212356567, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19957862794399261, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.19558092951774597, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.17882272601127625, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09852727502584457, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0942515954375267, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11010779440402985, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10157576948404312, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09973751753568649, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08963753283023834, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08571959286928177, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05586245283484459, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.048411279916763306, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04702269658446312, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04668733477592468, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.027852186933159828, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.023916788399219513, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.023800240829586983, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022085392847657204, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.021877184510231018, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014378750696778297, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014069161377847195, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013867196626961231, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00876921508461237, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09852727502584457, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09852727502584457, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.20070764422416687, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17225636541843414, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1627802699804306, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.13743449747562408, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09255282580852509, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08202590048313141, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10812371224164963, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09935262799263, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0962274968624115, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07352638989686966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06755322217941284, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.055833641439676285, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04813464730978012, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04507433995604515, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.044333070516586304, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02816169150173664, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024287888780236244, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02409830503165722, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.020701784640550613, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02022063545882702, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015703288838267326, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.016478372737765312, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014777756296098232, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012491429224610329, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.0962274968624115, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.0962274968624115, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.16274607181549072, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15409252047538757, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1513497531414032, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13874635100364685, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0762626901268959, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07322349399328232, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0844806656241417, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07808385044336319, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07715724408626556, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06956908851861954, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06664751470088959, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.042984385043382645, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03740053251385689, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03658318147063255, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.036387600004673004, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.021543940529227257, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.019021514803171158, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.018963389098644257, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.017689965665340424, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.0175694040954113, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011518903076648712, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011761657893657684, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011256814002990723, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008194755762815475, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0844806656241417, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0844806656241417, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.2102612853050232, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.19911229610443115, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19566091895103455, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17948582768440247, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09826688468456268, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09441490471363068, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.1087670549750328, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10056416690349579, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.0994219183921814, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08972015976905823, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08595846593379974, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05512066185474396, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04799248278141022, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.046976178884506226, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04673825949430466, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02755635417997837, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024018269032239914, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02394920215010643, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022254282608628273, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022107474505901337, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014372745528817177, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014232566580176353, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014039721339941025, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009162204340100288, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09826688468456268, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09826688468456268, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.21486712992191315, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18776312470436096, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.17645779252052307, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15590234100818634, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09735053777694702, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.0864371582865715, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11820030957460403, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.1079569160938263, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.10151595622301102, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08237231522798538, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07772058248519897, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.06002980098128319, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.051928140223026276, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04712913557887077, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.045931387692689896, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.0301984716206789, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.025166502222418785, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.024870865046977997, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02231927588582039, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.021567601710557938, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.016467276960611343, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.017042888328433037, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014943394809961319, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.012440570630133152, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09735053777694702, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09735053777694702, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10931004583835602, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10310453176498413, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10065588355064392, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09198111295700073, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.050837788730859756, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04847961291670799, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05746794119477272, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.052974965423345566, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05148952826857567, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04623159393668175, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.044382378458976746, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02916324883699417, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02530336007475853, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02434111014008522, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02410867251455784, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014595807529985905, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012525048106908798, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012430422939360142, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011598890647292137, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0114574721083045, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007651399355381727, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.00763256149366498, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007344461046159267, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005042601376771927, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09198111295700073, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09198111295700073, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09076282382011414, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08557683229446411, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08321462571620941, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07594752311706543, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04212942719459534, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03997955098748207, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04816214367747307, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04444754123687744, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0426662415266037, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.038300950080156326, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03670751675963402, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02438199520111084, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.021204901859164238, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.0201649721711874, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.01990606263279915, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01220188569277525, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010377416387200356, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010262355208396912, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009603800252079964, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009446432814002037, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006392897572368383, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006366898771375418, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.00606313394382596, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004178408533334732, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09076282382011414, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09076282382011414, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2199821174144745, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20786672830581665, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20392563939094543, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1865231692790985, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10270968824625015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09833940863609314, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11452724039554596, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10557298362255096, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10397861897945404, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09351275861263275, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08944976329803467, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.058086130768060684, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05033065751194954, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04900599271059036, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.048698484897613525, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028970981016755104, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024890966713428497, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.024787846952676773, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02298639714717865, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02279008738696575, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014921987429261208, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014563196338713169, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014423148706555367, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009008393622934818, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09351275861263275, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09351275861263275, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1787072718143463, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.15638841688632965, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1489136815071106, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.12163861840963364, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08260445296764374, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.07451659440994263, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09538798034191132, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08765044808387756, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08520381897687912, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06575768440961838, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0579984150826931, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.049203503876924515, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04283905774354935, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04054940491914749, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04000809043645859, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02498452551662922, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.022476665675640106, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.022339418530464172, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01945776678621769, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01910143345594406, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014354793354868889, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01587551087141037, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.01368042640388012, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012744766660034657, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09538798034191132, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09538798034191132, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.15615960955619812, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14790359139442444, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1453324854373932, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13335531949996948, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0731135904788971, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07023124396800995, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08085592836141586, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07483084499835968, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07397682219743729, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06674166768789291, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06396087259054184, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04115145280957222, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03579670935869217, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.035020679235458374, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.034840378910303116, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.020565304905176163, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.018111560493707657, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.018057215958833694, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.016834165900945663, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016721956431865692, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010872181504964828, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011061929166316986, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.01061999797821045, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007548756897449493, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08085592836141586, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08085592836141586, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20209430158138275, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.19146820902824402, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.18821023404598236, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1727280616760254, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09448280185461044, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09078052639961243, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10431528836488724, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09662296622991562, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.0955708846449852, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08627165853977203, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08255867660045624, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05291105806827545, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04606432840228081, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.0450899638235569, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04485731199383736, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02639920823276043, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.022938275709748268, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.022873464971780777, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.021250955760478973, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.021107181906700134, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.013647720217704773, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.013424273580312729, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.013324376195669174, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00840629544109106, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09662296622991562, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09662296622991562, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.2056856006383896, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.17864461243152618, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.16721756756305695, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.1478973776102066, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.0928153470158577, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08182572573423386, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11303344368934631, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10353391617536545, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09695331752300262, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.07815125584602356, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07409176975488663, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05755077674984932, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.04968133941292763, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04485975578427315, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04366535320878029, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.028976518660783768, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.023852339014410973, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.023546718060970306, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02104254812002182, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.02026974968612194, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015863453969359398, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016063891351222992, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014344677329063416, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011543691158294678, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09695331752300262, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09695331752300262, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.22.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11228799819946289, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10622309893369675, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1039181798696518, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09527017921209335, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05232508108019829, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05002247914671898, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05862680450081825, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05427255854010582, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05296922102570534, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.047753188759088516, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04570409655570984, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.029711632058024406, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025875521823763847, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.025000980123877525, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.024792999029159546, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014845442958176136, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01277250424027443, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012689625844359398, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011833831667900085, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01170072890818119, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.00771836144849658, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007620253134518862, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007431691512465477, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004858297295868397, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09527017921209335, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09527017921209335, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09766282886266708, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09232381731271744, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09017731249332428, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08265320956707001, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04547824710607529, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04340401664376259, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05134674161672592, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.047393668442964554, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.046028021723032, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04149697721004486, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03982091322541237, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02602720633149147, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022589052096009254, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02173297293484211, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.021529672667384148, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012996233999729156, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011097513139247894, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011016163975000381, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010282582603394985, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010153274983167648, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006745464634150267, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006631383206695318, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006466081365942955, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004207987803965807, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09766282886266708, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09766282886266708, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21793095767498016, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2063416987657547, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20264331996440887, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18600422143936157, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10187771171331406, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09778033196926117, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11325797438621521, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10448131710290909, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10312989354133606, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09299974888563156, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08904355764389038, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05739717185497284, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.049798384308815, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04861070588231087, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.048320069909095764, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028612129390239716, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024681050330400467, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02458735927939415, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022837800905108452, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02265629544854164, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014740413054823875, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014392751269042492, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0142986960709095, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008882319554686546, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09299974888563156, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09299974888563156, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.17985393106937408, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1447107195854187, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.13330723345279694, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1061617061495781, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08197008073329926, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.06863053888082504, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09692249447107315, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08865924924612045, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08591122180223465, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05906200408935547, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05277447775006294, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.049736276268959045, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04288826510310173, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03986551612615585, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03912524878978729, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.025093400850892067, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.021490175276994705, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.021329279989004135, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01727251522243023, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.016767676919698715, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013995055109262466, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014641323126852512, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013083296827971935, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011088121682405472, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09692249447107315, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09692249447107315, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.22.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.16302484273910522, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15449687838554382, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.15187141299247742, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13939043879508972, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07635293155908585, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07337063550949097, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08440382778644562, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07811099290847778, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07723329961299896, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06975895166397095, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06686627864837646, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04296161234378815, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.0373680405318737, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.036570802330970764, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03638487309217453, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.021469732746481895, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01891527697443962, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01886148750782013, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01759578287601471, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.0174784567207098, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011356563307344913, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011553616262972355, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011095171794295311, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00788825936615467, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08440382778644562, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08440382778644562, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.22.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20754240453243256, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1967468559741974, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19352836906909943, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17768588662147522, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0970211774110794, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09328547865152359, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10740648210048676, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09921014308929443, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09815829992294312, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08872158825397491, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08510111272335052, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.054547879844903946, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.0473284088075161, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04634096473455429, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04610893502831459, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.027262669056653976, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.023645061999559402, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.023580078035593033, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.021936308592557907, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02178841270506382, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.0142900962382555, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.013947375118732452, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.013976210728287697, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008899502456188202, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09921014308929443, "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": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09815829992294312, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.22.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20675309002399445, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.17969346046447754, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.16773883998394012, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.14855092763900757, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09328567236661911, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08207204192876816, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11447875201702118, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10488925129175186, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09755270928144455, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.07883777469396591, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07477959245443344, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05813498795032501, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05035105720162392, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04508965462446213, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.043767739087343216, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02919379621744156, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.02393242157995701, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.023567629978060722, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02115701511502266, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.020311957225203514, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015791485086083412, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016139600425958633, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014128566719591618, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.01149553619325161, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09755270928144455, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09755270928144455, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10938285291194916, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10327863693237305, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10099909454584122, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09247171878814697, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.050945330411195755, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04860909655690193, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05718060955405235, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05295399948954582, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05156480520963669, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0463918074965477, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.044420767575502396, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.028992749750614166, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025264408439397812, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02434677816927433, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.024127520620822906, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014482044614851475, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012468858622014523, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01237883698195219, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011545372195541859, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011406968347728252, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007540375459939241, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0074995760805904865, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007240159437060356, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0048406352289021015, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09247171878814697, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09247171878814697, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09226753562688828, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08716731518507004, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08501831442117691, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07779945433139801, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04296920448541641, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.040886636823415756, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.048758696764707565, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.045011915266513824, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04350757226347923, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.039149198681116104, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03758798539638519, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.024712545797228813, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02147526480257511, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.020542221143841743, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.020320715382695198, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012354210950434208, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010515631176531315, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01042152289301157, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009740659967064857, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009601198136806488, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006432507187128067, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006345169618725777, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006134744733572006, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004068603273481131, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09226753562688828, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09226753562688828, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21690000593662262, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20517313480377197, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.2013593316078186, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1845449060201645, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10130453109741211, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09716971218585968, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11261694878339767, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10399258881807327, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10250299423933029, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09237407147884369, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08847137540578842, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05710870400071144, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04954301565885544, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.048317015171051025, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04803362488746643, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02847389131784439, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024524329230189323, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.0244322307407856, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022677984088659286, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022494062781333923, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014641465619206429, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014310960657894611, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014182871207594872, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008833708241581917, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09237407147884369, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09237407147884369, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.17842933535575867, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13962294161319733, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.12440015375614166, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10273705422878265, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08160574734210968, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.06536364555358887, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09921899437904358, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0911833643913269, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08500029146671295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05730404704809189, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05402306839823723, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.051006898283958435, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04411143809556961, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.039788197726011276, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03871656209230423, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02570200525224209, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.021508118137717247, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.021126730367541313, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.017061464488506317, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.016309959813952446, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014223257079720497, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014919937588274479, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012899849563837051, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011207709088921547, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09921899437904358, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09921899437904358, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1672157645225525, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15853466093540192, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.15591949224472046, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1431567221879959, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07831250876188278, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07530432939529419, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0865606889128685, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08008264750242233, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07923530787229538, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07159143686294556, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06864721328020096, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04405032843351364, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03829829394817352, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03750723600387573, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03731125593185425, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.022000448778271675, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01937403343617916, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.019320707768201828, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018017463386058807, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.017898451536893845, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011610245332121849, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011790302582085133, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011352784931659698, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008003195747733116, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0865606889128685, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0865606889128685, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20926806330680847, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.19846539199352264, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19521108269691467, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17919951677322388, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09784746915102005, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09406598657369614, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10804778337478638, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10000258684158325, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09894829988479614, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08946091681718826, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0857713520526886, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05485372617840767, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.047682151198387146, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.0467040054500103, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04646867513656616, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.027369946241378784, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.023765577003359795, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.0237063467502594, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02204185537993908, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.021891430020332336, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014224953949451447, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.013917628675699234, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.013901954516768456, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008743707090616226, "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": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09894829988479614, "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": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09784746915102005, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.23.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20800265669822693, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18140524625778198, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.16925710439682007, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15032801032066345, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.0939239040017128, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08273351192474365, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11552858352661133, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10598903894424438, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.0981333777308464, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.07975323498249054, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07559479027986526, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05872611701488495, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05086703971028328, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04546697810292244, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.044114887714385986, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02985820174217224, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.024279505014419556, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.023900629952549934, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02158178575336933, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.020729126408696175, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.016803352162241936, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016554435715079308, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.015229033306241035, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011985333636403084, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.0981333777308464, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.0981333777308464, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11003237962722778, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10404610633850098, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10167635232210159, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09317440539598465, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05126121640205383, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.048952989280223846, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05785071477293968, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05335262790322304, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.051908642053604126, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04676127806305885, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.044887468218803406, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02934275195002556, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025469942018389702, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02451738715171814, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.024293500930070877, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014688920229673386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012552131898701191, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012459862045943737, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01163594238460064, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01149082463234663, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007697666063904762, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007546290289610624, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0073890695348382, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004859167151153088, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09317440539598465, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09317440539598465, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09437437355518341, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08926123380661011, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08712834864854813, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07980042695999146, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04399354010820389, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04195266589522362, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04973749816417694, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.045906029641628265, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04453274607658386, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04012374207377434, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03850356489419937, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.025188928470015526, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.021898984909057617, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02102048508822918, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.020814407616853714, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01258505042642355, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010735959745943546, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010652309283614159, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009944068267941475, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.00981185957789421, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.00653118547052145, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006431185640394688, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006245499476790428, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004078679718077183, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09437437355518341, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09437437355518341, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21238009631633759, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20112784206867218, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.19741389155387878, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1810983270406723, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0992429330945015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09518849104642868, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11034461110830307, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10186081379652023, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10041280835866928, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09056652337312698, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08675486594438553, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.055941320955753326, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04854409024119377, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.047342415899038315, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04705065116286278, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.027879277244210243, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024015942588448524, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02392316795885563, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02223309315741062, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022049836814403534, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014340870082378387, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014002904295921326, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013889390975236893, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008611945435404778, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.0992429330945015, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.0992429330945015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.17985542118549347, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14446566998958588, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.13368868827819824, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11295244097709656, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08110911399126053, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.06822442263364792, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09585363417863846, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08661729097366333, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08455465734004974, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.059994179755449295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05459262430667877, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.048880934715270996, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04230343550443649, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03989461809396744, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.039326176047325134, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02530737593770027, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02212885394692421, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.022006411105394363, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01836201176047325, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.017991840839385986, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015015450306236744, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01562235876917839, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014352597296237946, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012515575625002384, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09585363417863846, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09585363417863846, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1733599454164505, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16439586877822876, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1616930514574051, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1484580785036087, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.0812096819281578, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07807106524705887, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.089793860912323, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.0830458328127861, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08215054869651794, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07423298805952072, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.07115068286657333, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.045679207891225815, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03970298543572426, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03886512666940689, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03867407888174057, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.022823607549071312, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.020014381036162376, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.019956324249505997, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018599729984998703, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.018478965386748314, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.0120152672752738, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.0120833246037364, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011746851727366447, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008078400045633316, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.089793860912323, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.089793860912323, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21394920349121094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20292407274246216, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19958078861236572, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1832851767539978, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10008013993501663, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09625834971666336, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11069027334451675, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10230418294668198, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10126334428787231, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09152519702911377, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0878007709980011, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05628221482038498, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.048826735466718674, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04782102257013321, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04758891090750694, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028149912133812904, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024417489767074585, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024351203814148903, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02266600914299488, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02251293510198593, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014770454727113247, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014429868198931217, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014447360299527645, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009248035028576851, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09152519702911377, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09152519702911377, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.24.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.21116487681865692, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.1850806027650833, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1733154058456421, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15423046052455902, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09532783925533295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08451243489980698, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11612246185541153, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10706750303506851, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09949170053005219, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08137647807598114, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07720876485109329, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05902881920337677, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.051180895417928696, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04588818550109863, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04456459730863571, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.029596129432320595, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.024013211950659752, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.023640101775527, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02126738801598549, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.020411673933267593, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015826648101210594, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.015765616670250893, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.01415202021598816, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.010720568709075451, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09532783925533295, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09532783925533295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11255364865064621, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1063883900642395, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10403995960950851, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0953294187784195, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0524488165974617, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.050099171698093414, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05893689766526222, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05448843538761139, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05312252789735794, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04783700034022331, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04584077000617981, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.029886571690440178, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025996150448918343, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.025082804262638092, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.024857211858034134, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01493808999657631, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012831556610763073, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012743178755044937, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011886253021657467, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011748451739549637, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007790497969835997, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0076975468546152115, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007494400721043348, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004952974617481232, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.0953294187784195, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.0953294187784195, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09571380913257599, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09052912890911102, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08835604786872864, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08094661682844162, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.044640570878982544, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0425511971116066, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0504973866045475, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0466015562415123, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04519094526767731, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.040674638003110886, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.039050549268722534, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02560635469853878, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022229153662919998, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02133374661207199, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02111777663230896, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012796781025826931, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010896715335547924, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010808205232024193, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010092243552207947, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009958659298717976, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006645892281085253, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006529463455080986, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.00635515246540308, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00414125295355916, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09571380913257599, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09571380913257599, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2178296148777008, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20628006756305695, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20249885320663452, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18576835095882416, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10179343819618225, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09767646342515945, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11308924853801727, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10439211130142212, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10300575196743011, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09285473078489304, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08896149694919586, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057307880371809006, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.049746204167604446, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04854821413755417, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04826635122299194, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02859129011631012, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024626391008496284, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.024540530517697334, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022793103009462357, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02261091209948063, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014681818895041943, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014335552230477333, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014240056276321411, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008814673870801926, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09285473078489304, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09285473078489304, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1862841248512268, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14595848321914673, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.13181310892105103, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10964397341012955, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0838521420955658, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.06616818904876709, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10152964293956757, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09316103905439377, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08912909030914307, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06182771921157837, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05534917116165161, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.052319180220365524, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04488762840628624, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04067438095808029, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.039624202996492386, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02637738734483719, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02161495015025139, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.021363696083426476, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.017424777150154114, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.016696473583579063, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014505750499665737, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014539700001478195, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013195646926760674, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.010468197055161, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09316103905439377, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09316103905439377, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1734616905450821, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16451004147529602, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1617535799741745, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1485673487186432, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.08133368194103241, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07816661149263382, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0899718850851059, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08312556147575378, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08225049078464508, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07433202862739563, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.07138542830944061, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04580031707882881, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.039767153561115265, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.0389416366815567, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03874393180012703, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.022931333631277084, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.0201119314879179, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02005961537361145, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018705444410443306, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.018582720309495926, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.012165260501205921, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.01223465520888567, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011905157007277012, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008296390064060688, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0899718850851059, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0899718850851059, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21538510918617249, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.204233318567276, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.200903058052063, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1844860166311264, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10080432146787643, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09690656512975693, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11145118623971939, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10302115976810455, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10195644944906235, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09218332916498184, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08846805989742279, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05664166435599327, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.049162548035383224, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04815492779016495, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04791867733001709, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028287231922149658, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02456788904964924, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.0245014987885952, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02279636822640896, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02265206165611744, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014821662567555904, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014494802802801132, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014495711773633957, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009247157722711563, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09218332916498184, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09218332916498184, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.25.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.2090761363506317, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.1839924156665802, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.171983003616333, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.1534576416015625, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09435317665338516, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08366701751947403, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11593969166278839, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10683445632457733, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09847424179315567, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08108455687761307, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07716045528650284, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.058790549635887146, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.051022350788116455, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04542476311326027, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04401800036430359, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.029409924522042274, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.023778006434440613, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.02334992215037346, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.021171528846025467, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.02026854269206524, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015617387369275093, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.015683429315686226, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.013822425156831741, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.010632476769387722, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09847424179315567, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09847424179315567, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11011451482772827, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1041412502527237, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10182630270719528, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09327217936515808, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05136900022625923, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04903096333146095, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05798390507698059, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05341881886124611, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.052000001072883606, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.046851277351379395, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.045005060732364655, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02938937023282051, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025487905368208885, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.024555549025535583, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02433348447084427, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014699187129735947, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012547663412988186, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01245599053800106, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011626065708696842, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01148406695574522, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007649767212569714, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007505353540182114, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.00735095702111721, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0047811539843678474, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09327217936515808, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09327217936515808, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.095609650015831, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09043442457914352, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08828391879796982, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08094900846481323, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04455849900841713, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04248719662427902, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.050726160407066345, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.046519022434949875, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.045109666883945465, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04063427448272705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.039175327867269516, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.025682710111141205, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02220464125275612, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.021316690370440483, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02110547013580799, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01283936481922865, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01090172491967678, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010816474445164204, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010103254579007626, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.00997046660631895, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0067025357857346535, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.00654565216973424, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0064155408181250095, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004180788528174162, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.095609650015831, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.095609650015831, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21799491345882416, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20643243193626404, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20277154445648193, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18593579530715942, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10202168673276901, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09791193902492523, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.1133551150560379, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1046215370297432, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10323082655668259, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09308718144893646, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08922839909791946, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0575454905629158, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04989313334226608, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.048702120780944824, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04840826988220215, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028708497062325478, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0247988011687994, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02471558004617691, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02296685427427292, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022785943001508713, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014832735992968082, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014582446776330471, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014389597810804844, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009177516214549541, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09308718144893646, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09308718144893646, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.16072776913642883, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13222166895866394, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.12088517844676971, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09382376074790955, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07219824939966202, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.06107538193464279, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09005936980247498, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08216984570026398, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0767407938838005, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05406501889228821, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04865482449531555, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0462694950401783, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04010472819209099, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03559935837984085, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03447781503200531, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.023343751206994057, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.019783586263656616, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01944953389465809, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.016568519175052643, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.015834074467420578, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013043141923844814, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014403346925973892, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.011648827232420444, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01131780631840229, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09382376074790955, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09382376074790955, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.26.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1747354120016098, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16566842794418335, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1629018485546112, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.14966154098510742, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.08191976696252823, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.0787554457783699, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09061028063297272, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08377723395824432, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08286365866661072, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07489161938428879, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.07183728367090225, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04609007388353348, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04006046801805496, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03922181576490402, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03902466595172882, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.023037219420075417, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02021973393857479, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.020161744207143784, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018801694735884666, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.018676018342375755, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.012166033498942852, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.012247754260897636, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011898567900061607, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008240878582000732, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09061028063297272, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09061028063297272, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21463921666145325, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20353958010673523, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.20014628767967224, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1838768571615219, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10044919699430466, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.0965891107916832, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11094260215759277, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10268670320510864, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10158804804086685, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09185582399368286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08808846771717072, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05631953105330467, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04897717013955116, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04796377569437027, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.047722749412059784, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02813386544585228, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024432791396975517, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024359479546546936, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022659899666905403, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022507814690470695, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014652919955551624, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014339576475322247, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014324551448225975, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009051023051142693, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09185582399368286, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09185582399368286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.26.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20751844346523285, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18281948566436768, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.17117373645305634, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15310373902320862, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09375769644975662, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08331438899040222, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11495821923017502, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10588094592094421, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09788593649864197, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08086714893579483, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07702947407960892, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05867131054401398, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.050813499838113785, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04528432339429855, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.043905407190322876, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02965504862368107, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.023968489840626717, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.02356291003525257, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.021466920152306557, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.02060168981552124, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.01615600660443306, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016101360321044922, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014415106736123562, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011345846578478813, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09788593649864197, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09788593649864197, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11786568909883499, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11149931699037552, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10913581401109695, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10005311667919159, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.055032309144735336, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.052593860775232315, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06155356019735336, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.056995145976543427, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05571029335260391, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05015791952610016, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.048019666224718094, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.031169690191745758, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.027172043919563293, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02628343738615513, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.026073046028614044, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.015560570172965527, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013394162990152836, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01331409066915512, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.012395299971103668, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012258780188858509, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.008051164448261261, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007939350791275501, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007755144499242306, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00499267503619194, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06155356019735336, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06155356019735336, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10045448690652847, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09500743448734283, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09292754530906677, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0851958766579628, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04692283645272255, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04480143263936043, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.052874062210321426, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04876244440674782, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0474834181368351, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.042785849422216415, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04110537841916084, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.026818403974175453, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023255689069628716, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.022421816363930702, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02222365140914917, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013400092720985413, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01143638789653778, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011356640607118607, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01058847364038229, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010467099025845528, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006949772592633963, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006805231794714928, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006678073201328516, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004291614517569542, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09500743448734283, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09500743448734283, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.22155573964118958, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2097293585538864, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20600415766239166, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1889602392911911, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10355997085571289, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09939375519752502, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11490520089864731, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10609523206949234, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10481736809015274, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09448616951704025, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09053473174571991, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05828915163874626, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05059047043323517, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04942553490400314, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0491420216858387, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.029063904657959938, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025094490498304367, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.025007931515574455, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.023214012384414673, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.023031990975141525, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014970444142818451, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014620971865952015, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014536887407302856, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009022901766002178, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09448616951704025, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09448616951704025, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1604814976453781, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12076672911643982, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10576843470335007, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08351428061723709, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07065775990486145, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05523718148469925, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08833979815244675, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08136647194623947, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07646544277667999, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04758191853761673, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04609682038426399, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04532903805375099, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.039089642465114594, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03423487767577171, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03297920525074005, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02275601029396057, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.018010398373007774, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01769336126744747, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01359931193292141, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012646695598959923, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.012216692790389061, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.012129615060985088, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.010680477134883404, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008326690644025803, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08351428061723709, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08351428061723709, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.17415283620357513, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16513584554195404, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.16231681406497955, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.14907580614089966, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.08166422694921494, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07849740982055664, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09032044559717178, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.0835328996181488, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08263933658599854, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07460352778434753, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.07147453725337982, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04594736546278, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03992834314703941, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03908572718501091, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03888916224241257, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.022954698652029037, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.020099088549613953, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.020044617354869843, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018665488809347153, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01854553446173668, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.012082464061677456, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.012100540101528168, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011805107817053795, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008044266141951084, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09032044559717178, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09032044559717178, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21563825011253357, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20440533757209778, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.2010377049446106, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.18471580743789673, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10095097124576569, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09701408445835114, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11144483834505081, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10318775475025177, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10211048275232315, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09222599864006042, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08836400508880615, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.056532520800828934, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04919853061437607, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.048173073679208755, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04793314263224602, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028201287612318993, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02448277734220028, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024409962818026543, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02268110029399395, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022535230964422226, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014577070251107216, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014279687777161598, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014239528216421604, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008874700404703617, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09222599864006042, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09222599864006042, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20613986253738403, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.1822204291820526, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.17072223126888275, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.1529548317193985, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09314365684986115, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08296424895524979, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.1142616719007492, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.105261892080307, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09708812832832336, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08051741868257523, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07675839960575104, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05797748267650604, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.050221603363752365, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04475928097963333, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.043387699872255325, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.029054883867502213, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.023270923644304276, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.0228511281311512, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02076948620378971, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.019885839894413948, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015444150194525719, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.015147213824093342, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.013677659444510937, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.01001589372754097, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09708812832832336, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09708812832832336, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11500171571969986, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10870981961488724, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10633431375026703, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09738145023584366, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.053720396012067795, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0512780137360096, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06032240390777588, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05579625815153122, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05439291149377823, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.048925239592790604, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04685915261507034, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.030591318383812904, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.026632210239768028, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.025678107514977455, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.025455009192228317, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.015295756049454212, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013140582479536533, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.013045870698988438, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01216048188507557, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012018119916319847, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007971322163939476, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007873313501477242, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007656880654394627, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005046588368713856, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09738145023584366, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09738145023584366, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09811221808195114, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09273948520421982, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09060945361852646, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0830099955201149, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.045786600559949875, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04367770999670029, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05175317823886871, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04770112410187721, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04636958986520767, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.041734643280506134, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04007900133728981, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.026245443150401115, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022768866270780563, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.021890675649046898, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02168775536119938, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01310755219310522, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011174886487424374, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01108737662434578, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010342088527977467, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010208813473582268, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006802293937653303, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006673273630440235, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006512761116027832, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00421920046210289, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09811221808195114, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09811221808195114, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21653145551681519, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20489516854286194, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20126351714134216, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18446648120880127, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10124382376670837, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09709487855434418, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11241813004016876, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10384999215602875, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10248100012540817, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09228989481925964, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08835254609584808, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057040564715862274, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04948742687702179, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.048291850835084915, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04800677299499512, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02842872031033039, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02448110468685627, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.024396119639277458, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022626951336860657, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022448357194662094, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01460213866084814, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014236515387892723, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014151956886053085, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00871091429144144, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09228989481925964, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09228989481925964, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.17844755947589874, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12985007464885712, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11056094616651535, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0859268456697464, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07593991607427597, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.053962431848049164, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10013021528720856, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09082837402820587, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08424700796604156, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05189960077404976, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.049541816115379333, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05107840150594711, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04406873509287834, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.037413161247968674, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03569044545292854, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.026018455624580383, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.020686767995357513, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.020272456109523773, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.016426879912614822, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.015239066444337368, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014781039208173752, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.015217368490993977, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012832773849368095, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01161517295986414, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.0859268456697464, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.0859268456697464, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.28.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.17610995471477509, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16687016189098358, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.16405321657657623, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.15062132477760315, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.08266989886760712, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07941503822803497, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09149908274412155, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08455058932304382, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.0836547464132309, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07550409436225891, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0723278820514679, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.046585582196712494, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.040480297058820724, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03962337598204613, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03941800445318222, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.023351257666945457, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.020465752109885216, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.020414043217897415, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.019023707136511803, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01890021376311779, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.012462342157959938, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.012462019920349121, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.012189729139208794, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008464756421744823, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09149908274412155, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09149908274412155, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.28.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.2181415557861328, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20673014223575592, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.20325183868408203, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1866769641637802, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10216431319713593, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09816544502973557, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11308880895376205, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10442689806222916, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10333704203367233, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09331204742193222, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08944784104824066, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05743478611111641, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04989077150821686, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.048849791288375854, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04860382899641991, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028774619102478027, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02497936226427555, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024913514032959938, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.023165632039308548, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.023016516119241714, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.015201801434159279, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014807204715907574, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014868094585835934, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009564233012497425, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09331204742193222, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09331204742193222, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.28.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.21012714505195618, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18704888224601746, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.17554430663585663, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15744870901107788, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09540887176990509, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08530136942863464, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11669677495956421, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10771185904741287, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09915737807750702, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.0830167829990387, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07913393527269363, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05934980884194374, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05168084800243378, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04611051827669144, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.0446956641972065, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.029855720698833466, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.02441723644733429, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.023960521444678307, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.022023610770702362, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.02114925906062126, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.01595679670572281, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016427237540483475, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014134524390101433, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011603656224906445, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09915737807750702, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09915737807750702, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11793975532054901, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11153124272823334, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10919149965047836, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10011123865842819, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05512218549847603, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.052715349942445755, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.061658963561058044, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0570378378033638, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05580902099609375, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05022929981350899, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04805196821689606, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.031221779063344002, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.027205130085349083, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02633175067603588, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02612132951617241, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.015581916086375713, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013402813114225864, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.013322954997420311, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.012392289936542511, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012258928269147873, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.008046228438615799, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007918931543827057, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007753422949463129, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004948618821799755, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.061658963561058044, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.061658963561058044, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10135896503925323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09587405622005463, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.0938144102692604, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08597788959741592, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.047383151948451996, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04526560381054878, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.053300052881240845, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.049160849303007126, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04797810688614845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04319600388407707, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04145844280719757, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.027028493583202362, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023448100313544273, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.022639261558651924, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02245190553367138, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01349589228630066, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011542645283043385, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011464041657745838, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010680903680622578, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010559328831732273, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006989935413002968, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006851206999272108, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0067238821648061275, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004304597154259682, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09587405622005463, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09587405622005463, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.22189518809318542, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21001003682613373, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20621120929718018, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18914788961410522, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10382086038589478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09963622689247131, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11520877480506897, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10642808675765991, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10508324950933456, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09467575699090958, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09066296368837357, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05840971693396568, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05071956291794777, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.0495489276945591, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.049266815185546875, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02912597917020321, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025124792009592056, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02502969093620777, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02321173995733261, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.023042097687721252, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014958102256059647, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014592429623007774, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014529163017868996, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008926346898078918, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09467575699090958, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09467575699090958, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1894965022802353, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14134351909160614, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11939951777458191, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08957362920045853, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07851200550794601, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.061673350632190704, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10988163948059082, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10110054910182953, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09095556288957596, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.057044751942157745, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05512358248233795, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.056717902421951294, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04879981279373169, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.038560375571250916, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.035771310329437256, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028546664863824844, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02112090028822422, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.020406663417816162, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.017252011224627495, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.015335865318775177, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015721524134278297, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.015907449647784233, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012623891234397888, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011518179439008236, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08957362920045853, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08957362920045853, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.29.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.17538703978061676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16615283489227295, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.16336707770824432, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.14994044601917267, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.08230329304933548, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07907212525606155, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09119199216365814, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.0842263475060463, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08331353217363358, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07518425583839417, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.07207853347063065, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.046406339854002, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.040301594883203506, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.039445776492357254, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.039237093180418015, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02322085201740265, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02037413977086544, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02031385898590088, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018926473334431648, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.018801946192979813, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.012297512032091618, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.01239862758666277, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.012023111805319786, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008411495946347713, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09119199216365814, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09119199216365814, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.29.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21796664595603943, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20653481781482697, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.20308062434196472, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1864117532968521, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10210561007261276, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09807965159416199, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11276090145111084, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10440681129693985, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10329088568687439, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0932072177529335, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08922707289457321, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.057267092168331146, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04980438947677612, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04875757917761803, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04850657656788826, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028576785698533058, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024818839505314827, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024748245254158974, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02298552356660366, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02282860316336155, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014825972728431225, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.01454920694231987, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014479140751063824, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009147174656391144, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0932072177529335, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0932072177529335, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.29.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20635148882865906, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18411536514759064, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.17226125299930573, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15471699833869934, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09379804879426956, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08370690792798996, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11679264903068542, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10689900815486908, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09744027256965637, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08187223970890045, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07834692299365997, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.059595271944999695, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05133509635925293, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04542400315403938, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04391738772392273, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.03035767190158367, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.0241894219070673, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.02367589622735977, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02190111018717289, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.02097315527498722, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.016979357227683067, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016505969688296318, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.015147758647799492, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011796546168625355, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09744027256965637, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09744027256965637, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.12154791504144669, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11495906859636307, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11256986111402512, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10312876105308533, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05683479830622673, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05432494729757309, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06367393583059311, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05887101590633392, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05753466114401817, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0517766997218132, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04958697408437729, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03227908909320831, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02808847837150097, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02716873586177826, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.026942530646920204, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01613173820078373, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013865828514099121, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01378483884036541, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.012824656441807747, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01269221119582653, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.008379285223782063, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.008247551508247852, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.00806852150708437, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005232058931142092, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06367393583059311, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06367393583059311, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10433050990104675, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09871260821819305, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09656549245119095, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08848430961370468, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04883420094847679, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04664397984743118, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.054998330771923065, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05063263699412346, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04943731054663658, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04448496922850609, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04273618012666702, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.027918722480535507, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.024173030629754066, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02334832027554512, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02315015159547329, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013948392122983932, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011916397139430046, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011837848462164402, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.011025236919522285, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010901075787842274, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0072441259399056435, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.00708972429856658, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006968388799577951, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004490128252655268, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09871260821819305, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09871260821819305, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.22419145703315735, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21228237450122833, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.2084636688232422, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.19119799137115479, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10495687276124954, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.10067611187696457, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11634103953838348, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10753271728754044, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10620676726102829, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09568275511264801, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09159150719642639, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0590478777885437, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05125004053115845, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.05007774010300636, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.049799513071775436, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.029429439455270767, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025420252233743668, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.025330061092972755, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.023480692878365517, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02331230603158474, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015136667527258396, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014793218113481998, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014703357592225075, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009120174683630466, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09568275511264801, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09568275511264801, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2031799703836441, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14495471119880676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11817057430744171, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10191835463047028, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09031282365322113, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.06510463356971741, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11799558997154236, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10811066627502441, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09728939831256866, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0592329315841198, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.059728123247623444, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06070476397871971, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05202723667025566, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04395219683647156, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.041844431310892105, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.030448561534285545, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02353965863585472, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.0228132214397192, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.017801273614168167, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.016214076429605484, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.016437551006674767, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.016611572355031967, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013866530731320381, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011780006811022758, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.09728939831256866, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.09728939831256866, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.30.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.17110221087932587, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16200704872608185, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.15921816229820251, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.14612117409706116, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.08032849431037903, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07713575661182404, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08891232311725616, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08218704164028168, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08130136132240295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.0733209177851677, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.07028011232614517, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.0453157052397728, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03933129459619522, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03847966343164444, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.038278721272945404, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02265464887022972, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.019852928817272186, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.019795840606093407, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018432769924402237, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.018305731937289238, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011994917877018452, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.012047914788126945, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.01171975489705801, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00813197810202837, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08891232311725616, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08891232311725616, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21724016964435577, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20587797462940216, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.2022775411605835, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.18567973375320435, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10183234512805939, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09781647473573685, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11261487752199173, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10414787381887436, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10302748531103134, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09293702244758606, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08901841938495636, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05730544403195381, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04971470683813095, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04865964129567146, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.048403482884168625, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028598416596651077, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02483627013862133, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024767935276031494, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02301250584423542, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022856947034597397, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014954476617276669, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014669664204120636, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014611287042498589, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009376863949000835, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09293702244758606, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09293702244758606, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.30.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20842328667640686, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18598438799381256, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1744428277015686, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15684767067432404, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09463251382112503, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.0846870020031929, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11643271893262863, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10732667148113251, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09837382286787033, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08273229748010635, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07894663512706757, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05906257405877113, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05143993720412254, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04570290446281433, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04425595700740814, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02971351146697998, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.024139124900102615, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.02365013398230076, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.021824564784765244, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.020918823778629303, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015985317528247833, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.01619335077702999, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014182856306433678, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011302165687084198, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09837382286787033, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09837382286787033, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11793659627437592, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11144911497831345, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10907618701457977, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09993197023868561, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.055176761001348495, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05268878489732742, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.061772074550390244, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05718935653567314, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05586114525794983, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.050217706710100174, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04802633821964264, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03130185976624489, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02728486806154251, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.026373891159892082, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.026158269494771957, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01563095860183239, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013446059077978134, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.013360528275370598, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.012426414526998997, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012288440950214863, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.00809777993708849, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007983471266925335, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.00779732083901763, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005033255089074373, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.061772074550390244, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.061772074550390244, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10037173330783844, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09487645328044891, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09271649271249771, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.084966741502285, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04695672541856766, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04480830952525139, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05289780721068382, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04876290634274483, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04752989113330841, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04273808002471924, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04100320488214493, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.026814626529812813, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023266220465302467, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02244354784488678, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.022248266264796257, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013396930880844593, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011461466550827026, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011383388191461563, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010598527267575264, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010474140755832195, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006955267395824194, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006836839951574802, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006685921922326088, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004342273809015751, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09487645328044891, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09487645328044891, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.22499021887779236, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2128208875656128, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20910049974918365, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1915585696697235, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10530637949705124, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.10103567689657211, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11681096255779266, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10791447013616562, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10658088326454163, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09593971073627472, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09177134931087494, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.059204477816820145, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05142705887556076, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.05023990944027901, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.049973439425230026, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.0295250304043293, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02547011896967888, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02538534440100193, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02352507784962654, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02333953231573105, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01515375729650259, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014782516285777092, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014712593518197536, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009030360728502274, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09593971073627472, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09593971073627472, "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": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.17029014229774475, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12743480503559113, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10659286379814148, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09167058020830154, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07788557559251785, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05765055865049362, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10090327262878418, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09227866679430008, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08165278285741806, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.053052376955747604, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05188813805580139, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05211248621344566, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04451656714081764, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.0378744974732399, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.036155253648757935, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.026043448597192764, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.020213373005390167, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01948332041501999, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01557066012173891, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.014296136796474457, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013970572501420975, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014131823554635048, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.011801091022789478, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0099505465477705, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09167058020830154, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09167058020830154, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.16927602887153625, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16023577749729156, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1574048548936844, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.14438919723033905, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07947374880313873, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07626327872276306, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08790329843759537, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08134017139673233, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08042310178279877, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07245205342769623, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06941192597150803, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04476528242230415, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03891805186867714, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03806639462709427, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03786986693739891, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02239464409649372, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01964004524052143, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01957913301885128, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018219636753201485, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01809334009885788, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011849002912640572, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011918338015675545, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011576238088309765, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008025670424103737, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08790329843759537, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08790329843759537, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21694879233837128, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20534199476242065, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.20182543992996216, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.185195654630661, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.10162142664194107, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09755197912454605, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11225265264511108, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10392741858959198, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10282837599515915, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09266520291566849, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0887594074010849, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05699029564857483, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.049577970057725906, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04852746054530144, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.048275455832481384, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.0284267570823431, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024690495803952217, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024618547409772873, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022846782580018044, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022685930132865906, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014717545360326767, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014462657272815704, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014365948736667633, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009060214273631573, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09266520291566849, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09266520291566849, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.21027907729148865, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18722911179065704, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.17524167895317078, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.1574963629245758, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09540693461894989, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08516617119312286, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.1199018806219101, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10870768129825592, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.0993814766407013, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08338049799203873, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07984166592359543, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.06063627079129219, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05225132033228874, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04619782790541649, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04468581825494766, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.030794493854045868, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.02462448738515377, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.024135468527674675, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.0223158597946167, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.021383604034781456, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.017170913517475128, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016814935952425003, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.015317259356379509, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.012056359089910984, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09540693461894989, "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": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09540693461894989, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.32.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1202622503042221, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11354935169219971, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11112025380134583, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1016562432050705, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05622155964374542, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05365579202771187, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06302572786808014, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.058313097804784775, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05691836401820183, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05113040655851364, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04891238361597061, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.031945958733558655, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02781873568892479, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02686319872736931, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.026640918105840683, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.015945062041282654, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013704737648367882, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.013616406358778477, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01265844888985157, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012520386837422848, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.008261175826191902, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.008156845346093178, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007940723560750484, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005154531914740801, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06302572786808014, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06302572786808014, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.32.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10108943283557892, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09551426023244858, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09332343190908432, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08542966097593307, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04725797474384308, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04507887363433838, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.053322020918130875, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04916727542877197, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0478573702275753, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.043015554547309875, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04127226024866104, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.027073295786976814, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023472296074032784, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.022602999582886696, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02239818125963211, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013526839204132557, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011537527665495872, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011458897031843662, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010664469562470913, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010536417365074158, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007025847677141428, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.00688534090295434, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006735892966389656, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004361124709248543, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09551426023244858, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09551426023244858, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.32.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.23068822920322418, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21804960072040558, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.21411876380443573, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.19617322087287903, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10799628496170044, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.10352786630392075, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11985154449939728, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11066242307424545, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10935402661561966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09832371771335602, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09403170645236969, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06078251451253891, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.052759792655706406, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.051532894372940063, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.05123379826545715, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.030313139781355858, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.026141595095396042, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.026054663583636284, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.02412513457238674, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02393493428826332, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015575285069644451, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01519656553864479, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.015113744884729385, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009324351325631142, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09832371771335602, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09832371771335602, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.32.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1913633793592453, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12944696843624115, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1014971137046814, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08679045736789703, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07916631549596786, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05281722918152809, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.1101779043674469, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10093595087528229, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09171248972415924, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05378390848636627, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0548790767788887, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05671513453125954, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.048476725816726685, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03846975788474083, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.035730503499507904, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.0283652376383543, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02041368931531906, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.019728638231754303, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.015628669410943985, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.013555304147303104, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015122650191187859, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014714266173541546, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.01188703067600727, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00980677641928196, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08679045736789703, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08679045736789703, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.32.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.16635587811470032, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15739424526691437, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.15463599562644958, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1418253779411316, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07811792939901352, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07495512068271637, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08640414476394653, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07995979487895966, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07905624806880951, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07118213176727295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06818263232707977, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04403907060623169, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03827129676938057, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03742261976003647, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.037227414548397064, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.022019358351826668, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.019302841275930405, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.019247014075517654, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.017901966348290443, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.017780989408493042, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011637536808848381, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011713419109582901, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011362891644239426, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007891332730650902, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08640414476394653, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08640414476394653, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.32.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.2143964022397995, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20292387902736664, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19948077201843262, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1829453706741333, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.100526362657547, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09647130966186523, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11112122982740402, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10281624644994736, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10172776877880096, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09164322167634964, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08770930022001266, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.056484855711460114, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04909146577119827, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04802756756544113, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.047784555703401566, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02818586491048336, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024488911032676697, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.024421781301498413, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02266097441315651, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.022503074258565903, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014654645696282387, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014419683255255222, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.014309810474514961, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009149251505732536, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09164322167634964, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09164322167634964, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.32.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20191708207130432, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.17979541420936584, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1678343564271927, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15104927122592926, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09145257622003555, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08143489062786102, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11469364911317825, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10503923892974854, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09536656737327576, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.0800614207983017, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07667803764343262, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05854823440313339, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.050318893045186996, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04419949650764465, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04264825955033302, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.029863391071558, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.02337302453815937, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.022841086611151695, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.021163931116461754, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.02018885873258114, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.01681128889322281, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.015808915719389915, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.014951863326132298, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011009045876562595, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09536656737327576, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09536656737327576, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.33.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11682181060314178, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11023920774459839, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10778719931840897, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09864966571331024, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.054640065878629684, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.052085764706134796, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06140793114900589, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.056788429617881775, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05533267557621002, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04968666285276413, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04753490537405014, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.031140249222517014, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.027115141972899437, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.026128031313419342, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02589072845876217, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.015557571314275265, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013333390466868877, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.013239040039479733, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.012316077947616577, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012167568318545818, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.008074101060628891, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007950973697006702, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007751166820526123, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005038065370172262, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09864966571331024, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09864966571331024, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.33.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10011850297451019, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09451743960380554, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09231596440076828, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0845244824886322, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.046791672706604004, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.044624295085668564, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05272465571761131, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04868439957499504, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04737980663776398, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04256676882505417, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04074929654598236, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.026750829070806503, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0232304185628891, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.022371819242835045, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.022168558090925217, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013362261466681957, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011424977332353592, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011345668695867062, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010555779561400414, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010425988584756851, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006930380593985319, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006825170014053583, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006650153547525406, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004328512120991945, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09451743960380554, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09451743960380554, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.33.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2211894392967224, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2091635912656784, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20532555878162384, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1880088597536087, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10357778519392014, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09925016760826111, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11490167677402496, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10621558129787445, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10485372692346573, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0942167118191719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09013983607292175, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0582456961274147, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05062488093972206, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04941361024975777, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0491165854036808, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.029055409133434296, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025042500346899033, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.024954427033662796, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.023108916357159615, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022924408316612244, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014901910908520222, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014537740498781204, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014458674937486649, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008864708244800568, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.0942167118191719, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.0942167118191719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.33.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.16950242221355438, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1303064525127411, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11412565410137177, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0949515625834465, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07620606571435928, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05838088318705559, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.09716694056987762, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08774864673614502, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08059593290090561, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.054849978536367416, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05061499774456024, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04967624694108963, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.042625099420547485, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03741069510579109, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03609349578619003, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.025237197056412697, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02061745896935463, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02014932967722416, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.016801876947283745, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0159012358635664, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01445698644965887, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014887670986354351, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.012951911427080631, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011447440832853317, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.0949515625834465, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.0949515625834465, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.33.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.16416329145431519, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15534895658493042, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1526087075471878, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1398642659187317, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07712270319461823, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07400494813919067, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0853482037782669, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07898862659931183, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.078075110912323, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07026953250169754, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06725838035345078, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04345850646495819, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03781323879957199, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03697725385427475, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03678526729345322, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.021724173799157143, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01911946013569832, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.019062424078583717, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01773616299033165, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.017613176256418228, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011470605619251728, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011668656952679157, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011196507140994072, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00794125348329544, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0853482037782669, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.0853482037782669, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.33.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.21231979131698608, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.20089922845363617, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19743652641773224, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.18100684881210327, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09958025068044662, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.09555725753307343, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.11026482284069061, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.10192037373781204, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.10080374032258987, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09075073152780533, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08686777204275131, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.056033529341220856, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.04868851974606514, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.047620244324207306, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04737035930156708, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.028012828901410103, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.024364963173866272, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.02429366298019886, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.022559158504009247, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02239939756691456, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014703945256769657, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.014476452022790909, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.0143570052459836, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.009361409582197666, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09075073152780533, "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": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.09075073152780533, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.33.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20935070514678955, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.18619467318058014, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1741783767938614, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.15668915212154388, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09489688277244568, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08458865433931351, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11848299205303192, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10828498005867004, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09894338250160217, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.08287128061056137, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07927293330430984, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.06008746847510338, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.05192607268691063, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.045855797827243805, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.0443044938147068, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.030620770528912544, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.024200787767767906, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.023687778040766716, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02186262607574463, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.020887810736894608, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.017081359401345253, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.016286570578813553, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.01522334385663271, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011303507722914219, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09894338250160217, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09894338250160217, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.34.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11749230325222015, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11086726188659668, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10836441069841385, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09902681410312653, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05493184179067612, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.0523722879588604, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06170668452978134, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05713238567113876, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05566300079226494, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.049914486706256866, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04766872525215149, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03126733750104904, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02725157141685486, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.026276150718331337, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.026035545393824577, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.015617034398019314, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013419527560472488, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.013326410204172134, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.012384404428303242, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012236203998327255, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.008098587393760681, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.008010376244783401, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.0077760531567037106, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005087672267109156, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09902681410312653, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09902681410312653, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.34.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09855134785175323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09294470399618149, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09072743356227875, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08297290652990341, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0460660457611084, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04386506229639053, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.052076924592256546, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04807324334979057, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04665331169962883, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0418597050011158, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.040108777582645416, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02643013373017311, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022962074726819992, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.022060928866267204, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.021837379783391953, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013212930411100388, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011279947124421597, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.011192787438631058, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.0104193901643157, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010283066891133785, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006874174810945988, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0067814369685947895, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006578714121133089, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004336236044764519, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09855134785175323, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09855134785175323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.34.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2265801578760147, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2140723168849945, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.21015219390392303, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.19229482114315033, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.1060635894536972, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.10164669156074524, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11776159703731537, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10884323716163635, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10742919147014618, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09648703783750534, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09221852570772171, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.059746503829956055, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05188627541065216, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.050629571080207825, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.05033711716532707, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02979133650660515, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025675203651189804, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02559027448296547, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.023688213899731636, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.023496191948652267, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015299740247428417, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014933968894183636, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014839718118309975, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00914960540831089, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09648703783750534, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09648703783750534, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.34.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.18105480074882507, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13580597937107086, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11811093240976334, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08513887226581573, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0779464840888977, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.062445398420095444, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10048103332519531, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09260839968919754, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0863247811794281, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.051441680639982224, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05005481839179993, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05135262757539749, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04438411444425583, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.0376591831445694, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03592662885785103, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.02566704899072647, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.019841795787215233, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01940028741955757, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.014820890501141548, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.013461494818329811, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.013612248934805393, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01360347494482994, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.01146277878433466, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009183628484606743, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08513887226581573, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08513887226581573, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.34.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.16316786408424377, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15432798862457275, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.15158596634864807, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13888771831989288, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07665926218032837, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07352214306592941, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08484406024217606, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07850395888090134, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.0775899589061737, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06978066265583038, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06679927557706833, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04325105622410774, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.037575993686914444, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03673562780022621, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.036537546664476395, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.021622739732265472, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.018975140526890755, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01892050914466381, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.017593860626220703, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.0174711924046278, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011444449424743652, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011560526676476002, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.011174147948622704, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00783778540790081, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08484406024217606, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08484406024217606, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.34.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20552611351013184, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.19444946944713593, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.19104212522506714, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.17514319717884064, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09640799462795258, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.0924999937415123, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.10684141516685486, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09868840128183365, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09757764637470245, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08781743794679642, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.08409419655799866, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.05435316264629364, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.047134753316640854, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.046097755432128906, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04585486650466919, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.027169205248355865, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.023548536002635956, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.023481344804167747, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.02178865112364292, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.02163039520382881, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.014287997037172318, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.013947805389761925, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.013957162387669086, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008954278193414211, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09868840128183365, "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": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09757764637470245, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.34.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20065169036388397, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.17684483528137207, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.16476097702980042, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.14813317358493805, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.0904071107506752, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.08007927983999252, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11282296478748322, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.1038675606250763, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09485340863466263, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.07866878807544708, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.0752195417881012, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05737309157848358, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.04975533112883568, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04367106035351753, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.042137276381254196, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02878675051033497, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.023052016273140907, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.02256643958389759, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.020760465413331985, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.019789423793554306, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015459515154361725, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.01554290484637022, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.013532448559999466, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.010746657848358154, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09485340863466263, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09485340863466263, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.35.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11371143907308578, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10729040205478668, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10487217456102371, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09584818035364151, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.053144194185733795, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05067310482263565, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.0597204752266407, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05530034378170967, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05384554713964462, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04827509820461273, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04607938230037689, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.030243415385484695, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02636376954615116, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.025410883128643036, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02517848089337349, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01510678231716156, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012983296066522598, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012894482351839542, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01197786908596754, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.011835998855531216, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007848573848605156, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007763047702610493, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007533694617450237, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004943619016557932, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09584818035364151, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09584818035364151, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.35.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09538260102272034, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08997885137796402, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08785329759120941, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08031676709651947, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0445667989552021, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04245726764202118, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05035132169723511, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04646981135010719, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04513852670788765, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04049288481473923, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03873911127448082, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02553943358361721, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02217855304479599, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02132253348827362, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.021115845069289207, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012763690203428268, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010907408781349659, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01082252524793148, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010071134194731712, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009942026808857918, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006648034788668156, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006546269170939922, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006369005888700485, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0041923862881958485, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09538260102272034, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09538260102272034, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.35.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.22382035851478577, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2114269882440567, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20758472383022308, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18990668654441833, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10485600680112839, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.10044397413730621, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.11634151637554169, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10753708332777023, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10617449134588242, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09536069631576538, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0910457894206047, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05901648849248886, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05127720162272453, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.05005113407969475, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.049754928797483444, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.029427170753479004, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025380199775099754, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.025298595428466797, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.023409364745020866, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.023224778473377228, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.015116528607904911, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014765039086341858, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014669216237962246, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009060673415660858, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09536069631576538, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09536069631576538, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.35.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.15513533353805542, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12471599876880646, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11368679255247116, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08882629126310349, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0695703998208046, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.05852367728948593, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08501805365085602, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07810787856578827, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07414178550243378, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0504385270178318, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.044935375452041626, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04365536570549011, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03754310682415962, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.033697932958602905, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03275218605995178, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.021832464262843132, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.017819678410887718, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.017565956339240074, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.014151283539831638, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.013461906462907791, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.011742009781301022, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.011924290098249912, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.010526532307267189, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008398297242820263, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08882629126310349, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08882629126310349, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.35.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.16128034889698029, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.15248951315879822, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.14974641799926758, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1371317058801651, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07573660463094711, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07260597497224808, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08373310416936874, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07756632566452026, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07666893303394318, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06891682744026184, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06592673063278198, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04262744262814522, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03709982708096504, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.036269661039114, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.036071114242076874, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02129523456096649, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.018686382099986076, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01863016001880169, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.017313217744231224, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01718837209045887, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011182290501892567, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011314764618873596, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010911152698099613, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007577983662486076, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08373310416936874, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08373310416936874, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.35.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.19867971539497375, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.187851220369339, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.18459516763687134, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1691301017999649, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.09316980838775635, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.08937900513410568, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.1030028685927391, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09538861364126205, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.09434150904417038, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.08482895791530609, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0810924619436264, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.052324745804071426, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.045522868633270264, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04451365768909454, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.04427012801170349, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.026106229051947594, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.02266947738826275, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.022596973925828934, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.020947547629475594, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.0207955464720726, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.013532960787415504, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.01330526638776064, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.013194450177252293, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008373286575078964, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09538861364126205, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.09538861364126205, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.35.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.20141732692718506, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.176096573472023, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.16400952637195587, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.14757882058620453, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.09044277667999268, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.07979738712310791, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.11229059845209122, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.10351691395044327, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.0952427014708519, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.07826381921768188, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.07481139153242111, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.05704952031373978, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.04948951676487923, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.043568212538957596, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.04208020865917206, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02857516147196293, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.022806033492088318, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.022378429770469666, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.02040443941950798, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.019447974860668182, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015167484991252422, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.015125960111618042, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.013275561854243279, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.010179916396737099, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.0952427014708519, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.0952427014708519, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.36.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10968165844678879, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10335960984230042, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.10087759792804718, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09207714349031448, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05121646448969841, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04872436076402664, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05779377371072769, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0534675195813179, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05189567059278488, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.046461306512355804, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04436364769935608, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02929081954061985, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025524815544486046, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.024501275271177292, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02426130510866642, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01464124396443367, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012533587403595448, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.012430242262780666, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01155674085021019, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0114043690264225, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007622975390404463, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0075296820141375065, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.007296787109225988, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004815553314983845, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09207714349031448, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.09207714349031448, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.36.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08982881903648376, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08465801179409027, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08248886466026306, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07533590495586395, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04193985462188721, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03985395282506943, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.047555938363075256, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04396195709705353, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.042489685118198395, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03803428262472153, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.036359284073114395, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02409266121685505, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.020979415625333786, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.0200657919049263, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0198491383343935, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.0120437853038311, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010270647704601288, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010180666111409664, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.00947247352451086, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.00933564081788063, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0062743257731199265, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006201549433171749, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005979975685477257, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0039712912403047085, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08982881903648376, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08982881903648376, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.36.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.21895568072795868, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2068440169095993, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.20292669534683228, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1856367141008377, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10257714241743088, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09824065864086151, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.1139519140124321, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10524898767471313, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10385952889919281, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09319296479225159, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08901611715555191, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05786559730768204, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05020495504140854, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.048975009471178055, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0486886166036129, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.028849298134446144, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02487996034324169, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.024789519608020782, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022933553904294968, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.022748952731490135, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014840810559689999, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014533869922161102, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.014380116946995258, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008986106142401695, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09319296479225159, "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": 3.52529296875, "total_bits": 92413440.0, "err": 0.09319296479225159, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.36.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.16338743269443512, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1286896914243698, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.11668233573436737, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08774901926517487, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07398243993520737, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.06082064285874367, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.08937732130289078, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0816820040345192, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07853318005800247, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05049530416727066, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04593408852815628, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04612496867775917, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.039397284388542175, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.03588263690471649, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.03502007573843002, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.0231713205575943, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.019037790596485138, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01883404701948166, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.014590000733733177, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.013946413062512875, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.012665324844419956, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.012730712071061134, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.011549771763384342, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00914230290800333, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08774901926517487, "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": 2.72529296875, "total_bits": 71441920.0, "err": 0.08774901926517487, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.36.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.15555188059806824, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1470431387424469, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.14438477158546448, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.13222888112068176, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07299099117517471, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.0699891671538353, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08078030496835709, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.07479973137378693, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07392275333404541, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06642689555883408, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06351131945848465, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.041122086346149445, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.035785797983407974, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03497050330042839, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03477524220943451, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02054763399064541, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.018042901530861855, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.017988484352827072, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01671120896935463, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016590893268585205, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010808050632476807, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.010958464816212654, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010536441579461098, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007382551208138466, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08078030496835709, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08078030496835709, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.36.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.18639469146728516, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.17632023990154266, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1731332689523697, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.158613920211792, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.08741806447505951, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.08384117484092712, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09668650478124619, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08953027427196503, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08852079510688782, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07955662161111832, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.07606395334005356, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.04918580502271652, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.042738452553749084, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.04177919775247574, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.0415554940700531, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.024530814960598946, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.021306945011019707, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.021243704482913017, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01969042979180813, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.019546449184417725, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.012742412276566029, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.012559764087200165, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.012421383522450924, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007973520085215569, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09668650478124619, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.09668650478124619, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.36.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.19325947761535645, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.16650916635990143, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1542171835899353, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.13856148719787598, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.08651604503393173, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.07535409927368164, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.10895223170518875, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.09939711540937424, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09158005565404892, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.07399915903806686, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.0708610862493515, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.055358611047267914, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.04777771234512329, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.041991665959358215, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.040519580245018005, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02791421301662922, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.022478746250271797, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.022105714306235313, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.020092125982046127, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.01917348988354206, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015228398144245148, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.01551063358783722, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.01337665319442749, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011203591711819172, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09158005565404892, "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": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.09158005565404892, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.37.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10252513736486435, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09653766453266144, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09411974996328354, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08584791421890259, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.047796327620744705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04544982314109802, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.054183121770620346, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05002325028181076, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04845038801431656, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.043320998549461365, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04144015163183212, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02749544195830822, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023882025852799416, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02288878709077835, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.022649163380265236, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01375301368534565, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01177106611430645, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01167597807943821, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010866325348615646, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010717494413256645, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.0072104427963495255, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.0071730706840753555, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006893835961818695, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004713558126240969, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09653766453266144, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09653766453266144, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.37.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08730025589466095, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08222358673810959, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08003387600183487, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07304582744836807, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04070276394486427, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.038626719266176224, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.046384889632463455, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04282047599554062, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04126325622200966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03691452741622925, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03532463684678078, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.023522483184933662, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0204508937895298, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.01950204186141491, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.019272325560450554, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.011773206293582916, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010017499327659607, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009917952120304108, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.009246230125427246, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.009104972705245018, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006169249303638935, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006104875355958939, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005861527752131224, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.003976980224251747, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08730025589466095, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08730025589466095, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.37.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2058769166469574, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19418098032474518, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.1904749572277069, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1740274578332901, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09633215516805649, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.09220106154680252, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10717160254716873, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09893505275249481, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0975862368941307, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08740309625864029, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08346892148256302, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05440530180931091, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04718344286084175, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.045994002372026443, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.045721232891082764, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.027139980345964432, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.023388927802443504, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.023297514766454697, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.021540142595767975, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.0213603712618351, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01400325819849968, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.013695667497813702, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013571781106293201, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008515740744769573, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09893505275249481, "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": 3.17529296875, "total_bits": 83238400.0, "err": 0.0975862368941307, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.37.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.12430981546640396, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10429191589355469, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.0967385396361351, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07766290009021759, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.055599890649318695, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.048447348177433014, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.06825002282857895, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06219504773616791, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05878687649965286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.043331604450941086, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03850825875997543, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.035108525305986404, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.030542291700839996, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02758040279150009, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.026845784857869148, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.017887961119413376, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01572897844016552, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.015513299033045769, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.013641881756484509, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01319509744644165, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.010396873578429222, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.011746244505047798, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.009544454514980316, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009639845229685307, "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": 2.62529296875, "total_bits": 68820480.0, "err": 0.0967385396361351, "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": 2.62529296875, "total_bits": 68820480.0, "err": 0.0967385396361351, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.37.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.14989162981510162, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.14164689183235168, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1390838325023651, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1273626983165741, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.07037986069917679, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06745091825723648, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07802017033100128, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.0721554309129715, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.07126619666814804, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.06404387205839157, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.06127099692821503, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.039787545800209045, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03456605598330498, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.033761415630578995, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03356749191880226, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.019897418096661568, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01748797483742237, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.017432549968361855, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01621376909315586, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.016094500198960304, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010560670867562294, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.010730928741395473, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.01029216218739748, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007359708659350872, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07802017033100128, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07802017033100128, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.37.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1718456894159317, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.16245771944522858, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.15954044461250305, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.14613720774650574, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.08061086386442184, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.07729020714759827, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08930261433124542, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.08262038230895996, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.08163680136203766, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.07335945963859558, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.07012216001749039, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.045496225357055664, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.0394744873046875, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.03856774792075157, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03835432603955269, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.02272060513496399, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01975855976343155, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.019695663824677467, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.018270740285515785, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.018136385828256607, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.011920135468244553, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011775045655667782, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.01161880325525999, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007643180899322033, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08930261433124542, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.08930261433124542, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.37.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.1872786432504654, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.16025295853614807, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.14860771596431732, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.13307248055934906, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.08388059586286545, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.0728568285703659, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.1048644483089447, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.09556329995393753, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.0888507142663002, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.07107866555452347, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.06793346256017685, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.053534891456365585, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.046011488884687424, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.04075731709599495, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.039435043931007385, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.02707912027835846, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.021880222484469414, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.02158404141664505, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.01945287175476551, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.018625453114509583, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.015030511654913425, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.015114943496882915, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.013398585841059685, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.011048738844692707, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.09556329995393753, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.09556329995393753, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.38.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10084512829780579, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09508729726076126, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09270919859409332, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08463221788406372, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04706941917538643, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04475826770067215, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.053575702011585236, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.049302004277706146, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04768642038106918, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04274732246994972, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0409986786544323, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.027212556451559067, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023547057062387466, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02255745232105255, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.022321421653032303, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.013627809472382069, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011590429581701756, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01149003580212593, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010709059424698353, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.01056140847504139, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.007160869427025318, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.007046807557344437, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006847567856311798, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004602499771863222, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09508729726076126, "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": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09508729726076126, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.38.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.08428309857845306, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07942312210798264, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.07731597125530243, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07054910063743591, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03932197391986847, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.037307608872652054, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.044856272637844086, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04142560809850693, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03982546925544739, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.035691116005182266, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03412549942731857, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.022761039435863495, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.019778357818722725, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.018832260742783546, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.018605682998895645, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.011392820626497269, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009684755466878414, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.009581935591995716, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.008947507478296757, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.008804058656096458, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005993909668177366, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.005927837919443846, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005689286161214113, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0038766758516430855, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08428309857845306, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.08428309857845306, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.38.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2112581878900528, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19949740171432495, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.19581687450408936, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1789049357175827, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09891235083341599, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.094776451587677, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10983734577894211, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10145606845617294, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10017897933721542, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08985088765621185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08579453825950623, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05575517565011978, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.048390671610832214, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.04723644629120827, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.04695406183600426, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.027813348919153214, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024024371057748795, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.023940976709127426, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.022155743092298508, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.02197417989373207, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.01432102918624878, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.014064976014196873, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013895904645323753, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008755930699408054, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09891235083341599, "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": 3.22529296875, "total_bits": 84549120.0, "err": 0.09891235083341599, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.38.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.12821564078330994, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1051337942481041, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.09641732275485992, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07646214216947556, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05643346533179283, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.04892534762620926, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.07156311720609665, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06516947597265244, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06123165041208267, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04415840655565262, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03951379656791687, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.036860596388578415, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.031737953424453735, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.027872225269675255, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.026882562786340714, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.01873154565691948, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01557854749262333, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.0153294513002038, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.013402746059000492, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.012786978855729103, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.010702399536967278, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.011471269652247429, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.009570632129907608, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009056787006556988, "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": 2.62529296875, "total_bits": 68820480.0, "err": 0.09641732275485992, "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": 2.62529296875, "total_bits": 68820480.0, "err": 0.09641732275485992, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.38.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.12601302564144135, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.11910001933574677, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.11687535047531128, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.10701297968626022, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.05939250811934471, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.056931380182504654, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.06582770496606827, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.06090116873383522, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.06015462055802345, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.054082706570625305, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.0518190898001194, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.033895500004291534, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.02973879687488079, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.029076308012008667, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.0289112851023674, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.017134064808487892, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.016032220795750618, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.015988633036613464, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.015060316771268845, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.014968471601605415, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.009722902439534664, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.011103973723948002, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.009527268819510937, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.008982490748167038, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.06582770496606827, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.06582770496606827, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.38.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.13690988719463348, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.12941554188728333, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.12714123725891113, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.1165168359875679, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.06460339576005936, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.06198440492153168, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07183763384819031, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.06621888279914856, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.06537964940071106, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.05885832756757736, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.056501634418964386, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.036943987011909485, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.03233608230948448, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.031625233590602875, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.03144156560301781, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.018721923232078552, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.01744943857192993, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.01740865595638752, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.016392961144447327, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.01629788428544998, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.010742701590061188, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.012098251841962337, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.010530014522373676, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.00979590229690075, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07183763384819031, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.07183763384819031, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.38.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.1339680552482605, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.11343907564878464, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.1045907661318779, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.09337688982486725, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.06008928641676903, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.05161655694246292, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.07666254043579102, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.0689309686422348, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.06367354094982147, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.050614017993211746, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.04860299080610275, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.03948139771819115, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.03375297039747238, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.02967066690325737, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.02864590659737587, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.020376399159431458, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.01665477640926838, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.01638495735824108, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.01495351456105709, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.01430572010576725, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.011932825669646263, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.012286819517612457, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.010668684728443623, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.00970462430268526, "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": 2.727144820601852, "total_bits": 193024256.0, "err": 0.09337688982486725, "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": 2.727144820601852, "total_bits": 193024256.0, "err": 0.09337688982486725, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.39.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09511786699295044, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08957377076148987, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.08721699565649033, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07972922921180725, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04429903253912926, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.042043160647153854, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05076080560684204, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.046507883816957474, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04485645517706871, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04020291194319725, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0386558473110199, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.025741735473275185, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022220665588974953, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.021223222836852074, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.020983746275305748, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.012888442724943161, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010918214917182922, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.010815955698490143, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.010088218376040459, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.00993815902620554, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.006783584598451853, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.006675377022475004, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.006468434818089008, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004388537257909775, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09511786699295044, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09511786699295044, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.39.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.0750732421875, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07077030837535858, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.0687917098402977, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06278829276561737, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03497912734746933, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.033121541142463684, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.04005051404237747, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03696548566222191, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.035417962819337845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.031749218702316284, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.030362948775291443, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.020300298929214478, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.01766413077712059, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.016755621880292892, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.01654020883142948, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.010168490000069141, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.008648840710520744, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.008550251834094524, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.007993996143341064, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.007857894524931908, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.005361553747206926, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.005343164782971144, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.005072380416095257, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0035423669032752514, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0750732421875, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.0750732421875, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.39.self_attn.v_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.19650444388389587, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1854664385318756, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.18200623989105225, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.16641832888126373, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09201011061668396, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.08807265013456345, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.10370533913373947, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09440185129642487, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09311676770448685, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08364070951938629, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08041615039110184, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.052782170474529266, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04519851878285408, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.044074222445487976, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.0438130721449852, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.026759590953588486, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.022691255435347557, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.02260727249085903, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.020978132262825966, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.020809629932045937, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.014057443477213383, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.01361307967454195, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.013615967705845833, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008888536132872105, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09440185129642487, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09440185129642487, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.39.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09508129954338074, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07169904559850693, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 68820480.0, "err": 0.06297892332077026, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.054742537438869476, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.042000290006399155, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.72529296875, "total_bits": 97656320.0, "err": 0.03148236870765686, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 79464320.0, "err": 0.05443858355283737, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.047452207654714584, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.045170050114393234, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.029937932267785072, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.029911011457443237, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02816566452383995, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023824229836463928, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.22529296875, "total_bits": 110763520.0, "err": 0.02144297957420349, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.32529296875, "total_bits": 113384960.00000001, "err": 0.02084997668862343, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 131893120.0, "err": 0.014794591814279556, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012898045592010021, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.32529296875, "total_bits": 139599360.0, "err": 0.01277661882340908, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.52529296875, "total_bits": 144842240.0, "err": 0.01097982469946146, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.72529296875, "total_bits": 150085120.0, "err": 0.010636444203555584, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 158107520.0, "err": 0.009024472907185555, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 160570880.0, "err": 0.010175573639571667, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.2313232421875, "total_bits": 163350400.0, "err": 0.008342741057276726, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008778514340519905, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09508129954338074, "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": 2.17529296875, "total_bits": 57024000.0, "err": 0.09508129954338074, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.39.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.11418098211288452, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.10786554217338562, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1059177815914154, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.09698093682527542, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.05342976748943329, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.0512244775891304, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.05928831174969673, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.054777853190898895, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.05410228669643402, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.048615921288728714, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.04651384800672531, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.030181603506207466, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.026217449456453323, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.02561500109732151, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.025470774620771408, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.015100272372364998, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.013247039169073105, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.013209574855864048, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01227830071002245, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.012191164307296276, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.007985667325556278, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.008104066364467144, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.007785068824887276, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.005530655849725008, "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": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.09698093682527542, "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": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.09698093682527542, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.39.mlp.up_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.10933145880699158, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751085069444446, "total_bits": 168107520.0, "err": 0.1033024713397026, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251085069444446, "total_bits": 185802240.0, "err": 0.1014356017112732, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.09292086213827133, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2251085069444443, "total_bits": 228269568.0, "err": 0.05151497945189476, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7251085069444443, "total_bits": 263659008.0, "err": 0.04939005896449089, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031277126736111, "total_bits": 214550400.0, "err": 0.05728374049067497, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251085069444446, "total_bits": 221191680.0, "err": 0.052787989377975464, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1751085069444445, "total_bits": 224730624.0, "err": 0.05215948075056076, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5251085069444446, "total_bits": 249503232.0, "err": 0.04689128324389458, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6625542534722224, "total_bits": 259231488.0, "err": 0.045008596032857895, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031277126736111, "total_bits": 285329280.0, "err": 0.029498856514692307, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125108506944445, "total_bits": 291970560.0, "err": 0.025621581822633743, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.225108506944444, "total_bits": 299048448.0, "err": 0.025051238015294075, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.325108506944445, "total_bits": 306126336.0, "err": 0.024917909875512123, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031277126736111, "total_bits": 356108160.0, "err": 0.014894858933985233, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.225108506944444, "total_bits": 369827328.0, "err": 0.013553440570831299, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.325108506944445, "total_bits": 376905216.0, "err": 0.013518111780285835, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.525108506944444, "total_bits": 391060992.0, "err": 0.01268367376178503, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.725108506944444, "total_bits": 405216768.0, "err": 0.012605966068804264, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031277126736111, "total_bits": 426887040.0, "err": 0.008387384004890919, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125108506944445, "total_bits": 433528320.0, "err": 0.009099158458411694, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.231277126736111, "total_bits": 441042816.0, "err": 0.008206639438867569, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125108506944445, "total_bits": 575086080.0, "err": 0.007103152573108673, "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": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.09292086213827133, "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": 2.7251085069444443, "total_bits": 192880128.0, "err": 0.09292086213827133, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.39.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.05779476836323738, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 168120576.0, "err": 0.0481707826256752, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.62529296875, "total_bits": 185815296.0, "err": 0.04350343346595764, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.727144820601852, "total_bits": 193024256.0, "err": 0.03878673166036606, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.227144820601852, "total_bits": 228413696.0, "err": 0.025886170566082, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.736404079861111, "total_bits": 264458495.99999997, "err": 0.021643955260515213, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313232421875, "total_bits": 214553664.0, "err": 0.03488492593169212, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 221204736.0, "err": 0.030772550031542778, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1715892650462965, "total_bits": 224481536.0, "err": 0.027686644345521927, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.523441116898148, "total_bits": 249385216.0, "err": 0.021724920719861984, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6644983362268517, "total_bits": 259369088.0, "err": 0.021041620522737503, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 285332544.0, "err": 0.018006131052970886, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 291983616.0, "err": 0.015326867811381817, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.227144820601852, "total_bits": 299192576.0, "err": 0.013157001696527004, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3289966724537035, "total_bits": 306401536.0, "err": 0.012582306750118732, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.0313232421875, "total_bits": 356111424.0, "err": 0.009335007518529892, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.227144820601852, "total_bits": 369971456.0, "err": 0.007889003492891788, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.310478153935185, "total_bits": 375869696.0, "err": 0.007750326301902533, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.523441116898148, "total_bits": 390942976.0, "err": 0.007217875681817532, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.727144820601852, "total_bits": 405360896.0, "err": 0.006909511052072048, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.0313232421875, "total_bits": 426890304.0, "err": 0.005554204806685448, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.12529296875, "total_bits": 433541376.0, "err": 0.006320900283753872, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.235026945891204, "total_bits": 441308224.0, "err": 0.004849758464843035, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 575099136.0, "err": 0.0053742025047540665, "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": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.05779476836323738, "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": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.05779476836323738, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } } ], "last_module_idx": 82, "base_perplexity": 6.762514064754917 }